Conference Program | ACISM 2026

Full Program

Time Zone: Turkey Time - Istanbul (GMT+03:00)

Venue: NeoTech Campus, Istanbul Nisantasi University

Address: Ağaoğlu 1453, Maslak, 34000, Sarıyer, Istanbul, Türkiye

Friday, October 16, 2026 Day-1

Opening Ceremony Online Time: 09:30 - 10:00. Location: Keynotes Virtual Room

Join Session

Prof. Dr. Ali Hamitoğlu

Department of Computer Engineering, Istanbul Nişantaşı University, Turkey

Biography:

Dr. Ali HAMİTOĞLU obtained his Master’s degree in Computer Engineering from Eastern Mediterranean University, North Cyprus, followed by a Ph.D. from Selçuk University, Turkey. He completed his Postdoctoral research at Istanbul University-Cerrahpaşa. He has held academic positions at Istanbul Aydın University and Istanbul Sabahattin Zaim University. Dr. Ali has an extensive record of more than 150 publications in prestigious international journals and conferences, reflecting his substantial contributions to the field. He has played key leadership roles as Conference Chair, Program Chair, and Technical Program Committee member across numerous international conferences in USA, Europe, and other regions. In addition, he has served as Conference Proceedings Editor for more than 15 international conferences, as well as an editor for several academic books and a reviewer for numerous peer-reviewed journals. His editorial experience further includes serving as a Guest Editor for several SCIE-indexed journals. His research interests encompass Data Mining, Machine and Deep Learning, Big Data Analytics, Artificial Intelligence, Digital Signal and Image Processing, and Adaptive Filters. Through his active participation in various research collaborations and projects, he continues to advance innovation and excellence in these domains.

Prof. Dr. Hasan Alpay Heperkan

Faculty of Engineering and Architecture, Istanbul Nişantaşı University, Turkey

Biography:

Prof. Dr. Hasan Alpay Heperkan is a distinguished academic and mechanical engineering expert currently serving as the Dean of the Faculty of Engineering and Architecture at Istanbul Nişantaşı University, Türkiye. He graduated from Istanbul Technical University in 1974 and subsequently received Fulbright and TÜBİTAK scholarships to pursue advanced studies in the United States. He earned his M.Sc. from Syracuse University in 1976 and Ph.D. from the University of California, Berkeley, in 1980. His professional career includes research appointments at Lawrence Berkeley Laboratory and Union Carbide Research Center, as well as a visiting professorship at the University of Karlsruhe, Germany, supported by the Alexander von Humboldt Foundation. He became a full professor at Yıldız Technical University in 1997, where he also served as Dean of the Faculty of Mechanical Engineering from 2003 to 2006. Throughout his academic and professional career, Prof. Heperkan has made significant contributions to thermal sciences, energy efficiency, heating, ventilation and air conditioning (HVAC), numerical modeling, simulation, and optimization. He has held academic and leadership positions at several universities, including serving as Dean of the Faculty of Engineering at Istanbul Aydın University from 2016 to 2022. His extensive research portfolio includes numerous scientific publications, books, conference papers, and national and international research and industrial projects. He has supervised postgraduate and doctoral researchers, received national and international recognition for his contributions, and actively participated in professional engineering organizations, including ASHRAE and several Turkish engineering associations.

Prof. Dr. Ayşegül Komsuoğlu Çıtıpıtıoğlu

Istanbul Nişantaşı University, Turkey

Biography:

Ayşegül Komsuoğlu Çıtıpıtıoğlu graduated from Istanbul University, Faculty of Economics, Department of International Relations in 1997. That same year, she began her graduate studies and was appointed as a research assistant in the Department of International Relations from which she had graduated. She then continued her work as a research assistant at the Faculty of Political Sciences. She completed her doctorate in 2003 at Istanbul University's Institute of Social Sciences with a thesis titled "Political Leadership and the Case of Süleyman Demirel." As a postdoctoral researcher, she worked at Emory University's Department of Middle Eastern Studies in the United States. During the 2009–2010 academic year, she served at the University of California's Oral History Unit and at the Institute of Slavic, East European, and Eurasian Studies. In 2009, she received the title of Associate Professor in the field of "Political Life and Institutions." Following her academic studies abroad, she returned to Istanbul University's Faculty of Political Sciences. In 2014, she was promoted to full Professor.

Prof. Dr. Ayşegül Komsuoğlu conducts research in the areas of political sociology, Turkish political life, political philosophy, public policy, and migration studies. She also has research on the relationship between art and politics. Her articles and books have been published in national and international academic journals indexed in leading indices, and she has presented papers at international academic conferences in her field. Particularly in the field of migration studies, she has prepared and managed projects aimed at public policy formation in collaboration with TÜBİTAK, EU Framework Programs, and other institutions and organizations. Prof. Dr. Komsuoğlu has also supervised more than twenty master's and doctoral theses.

Keynote Session Online Time: 10:00 - 10:30. Location: Keynotes Virtual Room

Join Session

Dr. Marjan Asadinia

California State University, Northridge, USA

Talk: Self-Healing Memory Systems

Biography:

Dr. Marjan Asadinia is an Assistant Professor in the Department of Computer Science at California State University, Northridge, where she leads the Modern Computing and Communication Architectures (MOCCA) Lab. Prior to joining CSUN, she held academic and research positions at Cal Poly Humboldt, Bradley University, Oregon State University, and the University of Arkansas. Her research has been supported by the National Science Foundation (NSF) and industry partners such as Caterpillar Inc. She has received several awards for her research contributions, including the Research, Scholarship, and Creative Activity (RSCA) Award in 2025, Best Paper Presentation Awards at CCWC 2025 and AIIoT 2025, the Outstanding Research Competition Award at Cal Poly Humboldt in 2023, and the Faculty Scholarship Award in 2022. She has also published a book in Elsevier’s Advances in Computers series on Durable Phase-Change Memory (PCM) technology, highlighting its potential as a promising alternative to Dynamic Random-Access Memory (DRAM).

Keynote Session Online Time: 10:30 - 11:00. Location: Keynotes Virtual Room

Join Session

Dr. Mustafa Savaş Ünsal

İstanbul Nişantaşı University, Türkiye

Talk: Integrating Blockchain and Stablecoins into Crowdfunding Platforms

Biography:

Dr. Savaş Ünsal is an academic, technology executive, and entrepreneur with more than three decades of experience in information technologies, telecommunications, and digital innovation. He holds degrees in Mathematics from Middle East Technical University, Computer Science from Boğaziçi University, Business Administration from Oklahoma City University, and Business Management from İstanbul Nişantaşı University.

He served as Technical Director at TURKCELL and General Manager of SUPERONLINE, contributing to Türkiye’s telecommunications and Internet infrastructure. In 2010, he founded crowdFON.com, one of Türkiye’s pioneering crowdfunding platforms. He currently teaches Artificial Intelligence, Entrepreneurship, and Mathematics at İstanbul Nişantaşı University. His ACISM 2026 talk explores how blockchain and stablecoins can improve the efficiency, security, transparency, and liquidity of crowdfunding platforms.

Keynote Session Online Time: 11:00 - 11:30. Location: Keynotes Virtual Room

Join Session

Dr. Hadeel Saadany

Birmingham City University, United Kingdom

Talk: Large Language Models: A Double-Edged Sword of Promise and Peril

Abstract:This presentation delves into the multifaceted world of large language models (LLMs), exploring their strengths and limitations across various tasks, with a particular focus on their conversational capabilities. It will begin with a concise overview of the evolution of these models, the datasets that underpin their training, and their integration into proprietary applications that generate substantial revenue. Through real-world examples, the presentation will highlight inherent biases within LLMs and examine the critical errors they can produce in scenarios where they serve as the primary tool. It will also present state-of-the-art solutions for LLM bias and retrieval limitations.

Biography:

Dr. Hadeel Saadany is a Senior Lecturer and Deputy Course Lead in the Computer Science Department at Birmingham City University, United Kingdom. Her research spans multidisciplinary fields focusing on the intersection of Natural Language Processing (NLP), Deep Learning, and Artificial Intelligence (AI), with a core focus on solving real-life industrial problems through UK and USA industrial partners.

She previously worked at the University of Surrey, where she acted as lead researcher on a research-industrial project between eBay and the Institute for People-Centred Artificial Intelligence (PAI), applying recent Large Language Model technology to improve the efficacy of eBay's search engine infrastructure. She also led an Innovate UK research-industrial project that applied NLP techniques to build a multi-modal user interface facilitating legal information acquisition for the UK Supreme Courts, a system currently in the application process to be published as an innovative patent by the UK Intellectual Property Office. Her wider data science work has applied AI to regulatory compliance checking, business analytics in the automotive industry, cyber security risks in automatic machine translation, and fake news detection, alongside NLP research in automatic sentiment transfer, machine translation evaluation, and digital humanities.

Panel Discussion Online Time: 11:30 - 12:15. Location: Keynotes Virtual Room

Join Session

Topic: Women Shaping AI: Innovation, Intelligence, and Digital Transformation

Introduction & Objectives:

Artificial Intelligence is transforming the way we work, learn, communicate, and make decisions across almost every sector. As AI continues to evolve from data-driven analytics to intelligent software systems and AI-enabled organizations, women researchers, educators, and technology professionals are playing an increasingly important role in shaping its direction. This panel brings together women academics from diverse backgrounds in data science, software development, and management information systems to discuss how AI is changing technology, education, organizations, and society, while highlighting the opportunities for women to contribute to and lead the next generation of AI innovation.

Moderator

Dr. Sibel BORAN

İstanbul Nişantaşı University, Türkiye

Panel Members

Dr. Elanur Türküz

Istanbul Nişantaşı University, Türkiye

Dr. Banu Baklan Şen

İstanbul Nişantaşı University, Türkiye
University of Applied Sciences Europe Amsterdam, Netherlands

Dr. Nesibe Manav

İstanbul Nişantaşı University, İstanbul, Türkiye

Break Online Time: 12:15 - 12:30. Location: Keynotes Virtual Room

Join Session

Technical Session - 1 Online Time: 12:30 - 14:30. Location: Virtual Room -1

Join Session

Session Chairs: Dr. Elanur Türküz, Dr. Mustafa Savaş Ünsal

12:30
12:45

Paper ID: 44

Parameter-Efficient Fine-Tuning of CLIP with Cross-Modal Attention for Image-Text Retrieval and fMRI-Based Brain Decoding

Mohamed Hajji; Hounaida Moalla; Naima Halouani; Andrei Doncescu

Presentation: In Person

Abstract: Language-supervised vision models such as CLIP produce representations that align more closely with high-level visual cortex than purely visual networks, making them natural candidates for brain-to-image retrieval. Yet two challenges remain open: (i) adapting large CLIP backbones efficiently without catastrophic forgetting, and (ii) mapping cortical fMRI signals into a fine-tuned CLIP space for exhaustive image retrieval at low compute cost. We address both with a two-phase framework. Phase 1 — CLIP-CMA: we fine-tune OpenCLIP ViT-L/14 on MS-COCO 2017 using Gated Residual Adapters and a bidirectional Cross-Modal Attention (CMA) module (8 heads), with only 16.1% of the 437.9 M parameters trainable. Phase 2 — NeuroClipBD (BD-v2): a residual-MLP maps 15.7 k cortical voxels (NSD subject 01, 1.8 mm) into the CLIP-CMA embedding space via a combined InfoNCE + MSE loss, trained on a single GPU. CLIP-CMA reaches I2T-R@1 = 71.94% on COCO 5K (+6.9 pts over the frozen baseline). BD-v2 achieves Recall@1 = 16.8%, R@5 = 40.6%, R@10 = 55.3% on the exhaustive 982-image NSD pool (chance = 0.10%), a strictly harder protocol than the batch-sampled evaluation (300 candidates, chance = 0.33%) used by MindEye1/2; a 168× improvement over random in 30 epochs. Results are reported for a single NSD subject (subj01); multi-subject generalization remains future work.

12:45
13:00

Paper ID: 89

Beyond Binary Detection: A Stacked Ensemble over Memory-Forensic Attributes for Evasive Malware Family Attribution

Mohammad Imtiyaz Gulbarga; Ruslan Isaev; Ruslan Amanov

Presentation: In Person

Abstract: Evasive malware defeats signature-based defences by mutating its on-disk form while leaving unavoidable traces in volatile memory. Most memory-forensic studies report the binary benign-versus-malware task, on which published detectors already exceed 99% accuracy and where the remaining headroom is negligible. Such a detector tells an incident responder that a host is compromised but not what to do next. This paper addresses the harder four-class problem of attributing a memory image to Benign, Ransomware, Spyware or Trojan, which is what drives the operational response. We contribute an evasion-oriented reading of the volatile-memory attribute space that maps each attribute group to the concealment behaviour it evidences, an automated batch extraction pipeline built over Volatility, and a two-layer stacking ensemble combining Naive Bayes, Random Forest and Decision Tree base learners under a Logistic Regression meta-learner. On the 11,720-sample held-out partition of CIC-MalMem-2022 the ensemble attains 87.1% weighted accuracy and 0.87 weighted F1, with near-perfect benign separation (99.97%) and family-level accuracy between 71.4% and 78.8%. The 0.32 percentage-point margin over the strongest single learner is small but statistically significant under a paired McNemar test with Holm-Bonferroni correction (p = 0.0091). We further show that applying resampling outside the cross-validation folds inflates the accuracy estimate by 4.35 percentage points, a protocol error we quantify and avoid, and we report leak-free cross-validation (0.8734 ± 0.0026) that agrees with the held-out test accuracy to within 0.0025.

13:00
13:15

Paper ID: 47

Multimodal Cross-Attention for Automated Assessment of Sleep Apnea: A Preliminary Multi-task Study on Real Polysomnographic Signals

Sameh Horri; Wiem Ghars; Hounaida Moalla; Naima Halouani; Sameh Msaad; Samy Kammoun

Presentation: In Person

Abstract: Obstructive Sleep Apnea (OSA) is a prevalent respiratory disorder affecting approximately 936 million adults worldwide. Although polysomnography (PSG) remains the gold standard for diagnosis, its dependence on expert manual scoring limits large-scale screening. Existing automated approaches predominantly rely on single-modality signals and single-task formulations, restricting their clinical utility. We propose a multi-branch, multi-task deep learning architecture integrating five physiological modalities extracted from PSG: respiratory airflow (Flow), oxygen saturation (SpO2 ), thoracic effort, abdominal effort, and pulse rate. Each signal is independently processed through a dedicated 1D Convolutional Neural Network (CNN) paired with a Multi-Head Attention (MHA) mechanism to capture local morphological features and global temporal dependencies within 30-second windows. A cross-modal attention fusion module integrates the five branch embeddings, enabling each modality to attend to all others before concatenation, ensuring robustness under missing-sensor conditions. The frame-work simultaneously estimates the Apnea-Hypopnea Index (AHI) via regression, classifies the apnea type (central, mixed, obstructive, unknown), and predicts OSA severity through a shared dense backbone. Evaluated on a private clinical PSG dataset of 91 patients, the model achieved a test accuracy of 99.60% for OSA severity classification and 99.70% for apnea type classification, with AUC-ROC = 1.00 across all classes. Peak validation accuracy reached 99.69% (severity) and 99.72% (apnea type) at epoch 50. AHI estimation yielded a Mean Absolute Error (MAE) of 0.1223 on the regression scale, confirming reliable continuous regression. The proposed CNN-Attention architecture demonstrates strong clinical potential as a scalable, holistic screening tool, jointly optimizing regression and multi-class classification in a single pipeline while remaining robust to incomplete sensor configurations.

13:15
13:30

Paper ID: 112

When Conformal Coverage Does Not Transfer: A Cross-Benchmark Audit for Network Intrusion Detection

Mohammed Ahmed; Sacipe Toker; Kerem Ay

Presentation: In Person

Abstract: Conformal prediction (CP) promises set-valued intrusion-detection outputs with a finite-sample coverage guarantee, but the guarantee rests on exchangeability between calibration and test data, which realistic IDS benchmarks routinely violate. We report a duplicate-controlled, five-seed audit of split CP and binary class-conditional CP (BCC-CP) around a frozen Random Forest on the full public releases of NSL-KDD, CICIoT2023 (7.8 million records), and Edge-IIoTset (20.9 million records), within-dataset and under bidirectional NSL-KDD ↔ CICIoT2023 transfer through a four-feature proxy representation. At the 90% target, split CP reached 92.6% and 93.6% coverage within CICIoT2023 and Edge-IIoTset but only 67.3% on NSL-KDD's official partition, and collapsed to 22.5% and 0.01% under transfer; because the proxy space is almost perfectly separable by benchmark (balanced accuracy 0.9998), the transfer results are read as a severe benchmark-mismatch stress test. Per-class analysis shows that CICIoT2023's 92.6% marginal coverage concealed only 1.2% benign coverage, which BCC-CP raised to 89.7% within-dataset but could not restore under partition or cross-benchmark shift. In a pre-registered held-out attack-family test (0.46% unseen prevalence), CP prediction-set size was a poor novelty score (AUROC 0.160; AUPRC 0.0037, below chance) relative to Isolation Forest, entropy, maximum-probability, and Mahalanobis baselines. Conformal coverage in IDS should be treated as an empirically audited, benchmark- and class-specific property, and set-size novelty claims should be made only alongside dedicated out-of-distribution baselines.

13:30
13:45

Paper ID: 116

Integrated SHM-Based Smart CFRP Systems for Structural Performance and Sustainability

HASAN EREN ERBAŞ; CAN BALKAYA

Presentation: In Person

Abstract: This study proposes an integrated assessment framework for SHM-based smart CFRP systems by jointly evaluating structural performance, monitoring capability, and life-cycle sustainability within a unified perspective. Unlike conventional approaches, the proposed framework integrates structural behavior, sensing technologies, digital intelligence, and sustainability assessment into a single evaluation structure. A comparative literature review is conducted to examine CFRP-based systems and representative SHM-integrated applications. Structural performance is assessed in terms of load-carrying capacity, stiffness, crack control, energy dissipation, and durability. SHM performance is evaluated through continuous monitoring capability and fiber-optic-based damage detection. The framework also considers artificial intelligence and machine learning as supporting tools for processing SHM data for damage identification, condition assessment, and predictive maintenance. Sustainability is evaluated using life-cycle indicators including energy, life-cycle cost, maintenance demand, and service life. The assessment indicates that CFRP systems can improve structural performance and durability, particularly under seismic and harsh environmental conditions. SHM integration enables continuous condition assessment and supports more targeted maintenance strategies, while AI-assisted analysis can further support monitoring-data interpretation and maintenance decision-making. Although CFRP materials involve relatively high initial embodied energy, their durability, extended service life, and reduced maintenance demand may improve overall life-cycle performance. Overall, the proposed framework demonstrates the potential of SHM-based smart CFRP systems to transition from passive strengthening solutions toward intelligent and sustainability-oriented infrastructure by integrating structural performance, monitoring technologies, data-driven decision-making, and life-cycle considerations.

13:45
14:00

Paper ID: 118

Listening Through the Same Ear: Infrastructural Isomorphism and the Dispossession of Market Knowledge

Selcen Ozturkcan; Mahdi Ashraf Gandomy; Altug Tanaltay

Presentation: In Person

Abstract: Organizations have never had access to more consumer expression and never had less standing to interpret it. As generative intermediaries increasingly stand between firms and the people they serve, market knowledge arrives already summarized, already normalized, and already shaped by a party with independent interests. This conceptual paper theorizes that condition as organizational epistemic dispossession, in which the firm loses not access to consumer expression but the standing to warrant its own claims about its market. Building on this account, the paper extends institutional theory by specifying a fourth mechanism of organizational homogenization. Coercive, mimetic and normative isomorphism each presuppose that firms respond to one another or to a shared field. Infrastructural isomorphism does not. Competing firms respond independently and in good faith to what each experience as private evidence about its own market, and converge regardless, because that evidence has a common source. Convergence occurs without imitation, below the level at which strategic differentiation is theorized, and remains invisible to the firms subject to it. The paper develops a conceptual framework linking dispossession, signal degradation, propagation into analytics practice, and legitimacy without evidentiary warrant, and states ten propositions specifying the relationships among these constructs. Boundary conditions are specified, and directions for empirical operationalization are set out.

14:00
14:15

Paper ID: 121

Flexible Definition of Tests in GLARE-Edu

Alessio Bottrighi; Luigi Castello; Antonio Maconi; Jason Secreto; Gianluca Troncone; Erica Raina

Presentation: In Person

Abstract: Since 2019, we work on GLARE-Edu, a system that manages computer-interpretable clinical guidelines to support medical education. In this work, we describe our approach to support the definition of tests, aimed at ascertaining the ability of learners to apply guidelines on case studies. We propose an expressive representation formalism for tests, supporting the possibility of focusing on specific parts of the guidelines, and of fixing specific levels of abstraction. We also introduce a flexible mechanism to support teachers in the incremental and flexible definition of tests.

14:15
14:30

Paper ID: 122

Toward an Ontology-Based Framework for Multi-Dimensional Clinical Action Assessment: Distinguishing Semantic and Operational Errors

Erica Raina

Presentation: In Person

Abstract: Assessing clinical actions in medical education requires not only determining whether a learner decision is correct, but also preserving and explaining different dimensions of discrepancy between learner and guideline-recommended actions. Traditional assessment approaches often rely on monolithic evaluations that provide limited insight into the nature of learner–guideline discrepancies. At the same time, ontologies provide rich and shared representations of clinical knowledge that have been widely used to support semantic comparison and similarity assessment. Existing ontology-based approaches can determine how different two clinical actions are, but do not explicitly distinguish the different dimensions underlying such differences. We present an ontology-based framework for multi-dimensional clinical action assessment. The framework supports the structured decomposition of learner–guideline discrepancies into distinct and interpretable components, focusing on semantic and operational aspects of clinical actions. Beyond scoring, the framework supports explainable feedback and allows instructors to configure the relative importance of each assessment component, providing a pedagogically controllable, flexible, and extensible assessment framework supporting different learning objectives and stages of instruction. The framework is integrated into GLARE-Edu, an intelligent tutoring system for guideline-based medical education, but is independent of the specific ontology and tutoring platform adopted.

Technical Session - 2 Online Time: 12:30 - 14:30. Location: Virtual Room - 2

Join Session

Session Chairs: Dr. Nesibe Manav

12:30
12:45

Paper ID: 59

A Static and Adaptive Filtering Framework for Improved Semantic Analysis of Individual Driving Styles

DANUT DRAGOS DAMIAN

Presentation: In Person

Abstract: Identifying a driver’s style is challenging because it involves technical skills, behavioural patterns, and contextual factors. This paper proposes a novel approach based on semantic representations that segments the time series driving data into windows with similar characteristics to enable explainability and cross-driver comparison. The ultimate goal is to demonstrate that a lightweight method can be developed at low computational cost while still delivering competitive results. This research examines whether static or adaptive filtering techniques applied to signals from inertial sensors, such as accelerometers and gyroscopes, effectively preserve the dynamic features essential for recognizing aggressive driving. Median, Gaussian, and Savitzky–Golay static filters, which encode baseline and driver-independent safety patterns, are examined. It also evaluates Kalman and Alpha-Beta adaptive filters, which recalibrate data based on observed personal tendencies. The ability of these filters to maintain meaningful, recurring signal patterns is assessed using Recurrence Plots (RPs) and Recurrence Quantification Analysis (RQA). Data from three sources, Mendeley, Kaggle, and a proprietary dataset, are analysed. Results show that the Kalman filter most effectively preserves behavioural signal dynamics, exhibiting the highest values for determinism (DET), laminarity (LAM), entropy (ENTR), and average recurrence line length. These metrics indicate a structured and predictable signal. Another validation method involved classification using MobileNetV2, EfficientNetB0, and ResNet-50 models and the RP images dataset of aggressive and normal driving styles. The classification results confirm that the Kalman filter is the most effective preprocessing method for IMU data in driving behaviour recognition tasks.

12:45
13:00

Paper ID: 58

A hybrid CNN-ML architecture for improved forest fire classification – a meta-learner approach

ROBERT LEPADATU

Presentation: In Person

Abstract: This study aims to improve the performance and efficiency of pre-trained CNN models (MobileNetV3, DenseNet121, InceptionV3, ConvNeXt) for forest fire detection and analysis, in a meta-learner framework. An optimized training framework that uses both Adam and SGD optimizers is implemented. To maintain the performance stability and prevent overfitting, models are trained for 5, 10, 15, and 20 epochs. The achieved results show that InceptionV3 and DenseNet121 perform well on all datasets and classification optimization approaches. InceptionV3 achieved an accuracy of 0.995 and an F1-score of 0.995. DenseNet121 demonstrated stability and balance across most tests. They also perform consistently better than ConvNeXt and MobileNetV3. The models’ performance is evaluated on two benchmark datasets: the Dataset for Forest Fire Detection (D1) and the Forest Fire Image Classification Dataset (D2). Data augmentation techniques are applied to enhance the training dataset and improve model generalization. To further improve classification, we propose a hybrid CNN-ML architecture. DenseNet121 and InceptionV3 convert complex visual data into robust, lower-dimensional numeric feature vectors. The final classification layers of the CNNs are removed and replaced with a Sigmoid activation function for binary classification. Forward-passing images through the network yields highly discriminative feature representations from the bottleneck layer. The pertinent retrieved feature vectors are evaluated by Logistic Regression, SVM, and XGBoost classifiers. A deep learning algorithm combined with various base classifiers functions as a meta-learner, ultimately producing an accurate and optimized classification. The best results were achieved using DenseNet121-based feature extraction combined with Logistic Regression and SVM, with accuracies of 0.997 on dataset D1 and 0.996 on dataset D2, respectively.

13:00
13:15

Paper ID: 151

A Multi-Agentic Architecture for AI-Assisted Teaching and Assessment

Zain Ul Abideen Tariq; Ahmed Sheikh; Mowafa Househ

Presentation: In Person

Abstract: Large language models (LLMs) are increasingly being integrated into educational workflows; however, most existing educational applications rely on a single-model interaction paradigm in which one model is expected to generate instructional content, answer questions, assess student work, and provide feedback. This approach can limit transparency, task specialization, verification, and instructor control, particularly when the same model generates and evaluates educational outputs. This paper proposes a coordinated multi-agent architecture for AI-assisted teaching and assessment, instantiated through the \textit{HikmaProf} framework. The architecture decomposes educational workflows across specialized agents responsible for instructional planning, content generation, learner interaction, assessment generation, response evaluation, feedback, verification, and workflow coordination. An orchestration layer manages task delegation, information exchange, validation checkpoints, and escalation to human evaluators when academic or pedagogical judgment is required. The framework is examined through architectural analysis and representative teaching and assessment use cases demonstrating how multiple agents can collaborate across the instructional lifecycle. The analysis focuses on functional separation, traceability, verification, and human oversight rather than autonomous replacement of educators. The principal contribution is a modular and extensible architecture for educational artificial intelligence that separates content generation from evaluation, enables cross-agent verification, preserves instructor authority, and provides a foundation for developing more accountable and trustworthy agentic systems for teaching and assessment.

13:15
13:30

Paper ID: 165

Fire Risk Prediction for Medium- and Low-Rise Residential Buildings in the UAE Using Machine Learning Methods⁠

Shamma Ahmed; Waqar Ahmed

Presentation: In Person

Abstract: Residential fire incidents continue to represent a serious life-safety challenge in the United Arab Emirates, where rapid urban development, diverse building stock, extreme cooling loads, and limited publicly available incident data complicate predictive analysis. This study presents a data-driven framework for predicting fire severity in medium- and low-rise residential buildings using a secondary dataset of 50 UAE incidents recorded between 2019 and 2022. The analytical framework integrates logistic regression, random forest, gradient boosting, support vector machine with a radial-basis-function kernel, XGBoost, multilayer perceptron, and a stacking ensemble. To compensate for unavailable maintenance and spatial records, the Infrastructure Decay Index and Estimated Density Proxy were introduced as domain-informed variables. In the dedicated SVM ablation study, the raw-data ROC-AUC of 0.687 increased to 0.851 after adding the Infrastructure Decay Index, whereas the Estimated Density Proxy alone produced limited improvement. Binary test results reached 0.933 accuracy and 1.000 recall for SVM; however, the test set contained only 15 cases. Multi-class accuracy was unstable, falling from approximately 0.74–0.78 in cross-validation to 0.20–0.33 on the test set. SHAP analysis identified infrastructure decay, building age, smoke-detector absence, and fire cause as influential signals. The results demonstrate the feasibility of the framework as a research prototype, while data leakage, small-sample instability, and the absence of external validation prevent operational deployment.

13:30
13:45

Paper ID: 35

From Localhost to Campus: Lessons Learned Developing an AI and RAG-powered Chatbot Farm

Claudia Martínez-Araneda; Mariella Gutiérrez; Pedro Gómez; Nicolás Cereceda; Alejandra Segura-Navarrete

Presentation: In Person

Abstract: This article describes the development of a virtual assistant farm powered by Generative AI (Gen AI) and Retrieval-Augmented Generation (RAG) at the Universidad Católica de la Santísima Concepción (UCSC), highlighting best practices and key lessons learned. It traces the evolution from using no-code platforms to implementing a proprietary infrastructure called a “chatbot farm”. This study describes and analyses how this architectural transition overcame critical design constraints, reduced complexity in rapid engineering, and achieved an in-house development rate that, through the use of specialised frameworks, not only allows the implementation of new functionalities but also enables precise management of the large language model, reducing latency and improving data security. The paper considers recommendations, including hardware and environment, input validation and security, large language model management, user interface and experience (UI/UX) design, the quality assurance cycle, and scalability and maintainability. Future work will use medium-sized open-source models to reduce reliance on private API costs and improve the poor performance observed with lightweight models.

13:45
14:00

Paper ID: 42

Dual-Expert Seasonal-Trend Forecasting with State Space Models, Multi-Scale Convolutions, and Adaptive Gating

Ilhem Ltifi; Ilyes Gouider; Hounaida Moalla; Naima Halouani; Andrei Doncescu

Presentation: In Person

Abstract: Precise long-term time-series forecasting remains a challenge due to non-stationary data distributions and the difficulty of capturing both long-range dependencies and local temporal continuities. In this paper, we propose HSTN (Hybrid Seasonal-Trend Network), a novel architecture designed to address these challenges through a dual-stream decomposition and expert-gating mechanism. HSTN first applies Reversible Instance Normalization (RevIN) to mitigate distribution shift, followed by a moving-average series decomposition that explicitly decouples the input into seasonal and trend components. To process these components effectively, we introduce the HSTN block, which employs a global-local expert strategy: a Selective State Space Model (SSM) captures global, long-term dependencies within the seasonal signal, while a multi-scale dilated depthwise-separable convolution module (MultiScaleConv) extracts local continuity and sudden changes within the trend. A Dynamic Trust Gate adaptively fuses these representations by computing sample-specific weights, allowing the model to prioritize the most relevant expert. Experiments on seven benchmark datasets show that HSTN offers a favourable balance between computational efficiency—leveraging parallelized selective scans with near-linear complexity in the sequence length—and predictive accuracy over prediction horizons of 96 to 720 steps. The complete modular implementation, datasets, and execution scripts are publicly available at https://github.com/ilyesgouider/HSTN-Time-Series-Forecasting.

Saturday, October 17, 2026 Day - 2

Technical Session - 5 Online Time: 09:00 - 11:15. Location: Virtual Room -1

Join Session

Session Chairs: Dr. Waleed M.Ead

09:00
09:15

Paper ID: 109

A Hybrid Deep Learning Framework for Stock Movement Prediction: Evidence from Borsa Istanbul

Alperen Ünaldı; Somaiyeh Dehghan

Presentation: Online

Abstract: Stock movement prediction (SMP) is an important research area that enables investors to make more informed trading decisions and improve investment returns. However, accurately predicting stock market movements remains challenging due to the nonlinear, dynamic, and highly volatile nature of financial markets. Traditional forecasting approaches that rely solely on historical market data may fail to capture the impact of investor sentiment reflected in financial news. To address this limitation, this study proposes a hybrid deep learning framework that integrates historical stock market data and technical indicator with financial news sentiment for stock movement prediction. The proposed framework is evaluated using data from four companies listed on Borsa Istanbul (BIST). Financial news articles related to these companies are analyzed using a FinBERTurk sentiment analysis model adapted to the Turkish financial domain. Both daily sentiment and its three-day moving average are then fused with historical market data and technical indicators and provided to an LSTM network for stock movement prediction. Results show that the feature configuration combining daily sentiment and its three-day moving average achieves the best performance across all four companies, improving the average accuracy and macro-F1 by 2.56 and 2.72 percentage points, respectively, over the baseline. These findings demonstrate the effectiveness of incorporating short-term financial sentiment trends, particularly the three-day moving average, for improving stock movement prediction in the Turkish stock market.

09:15
09:30

Paper ID: 124

BCK-GAN with Grad-CAM for Alzheimer's Disease Detection from MRI Images

Adhwaith Ashok; Anjana S; Amanya M

Presentation: Online

Abstract: Alzheimer's disease (AD) causes progressive cognitive decline, and early diagnosis remains a persistent clinical challenge. MRI-based computer-aided diagnosis has shown promise, but many deep learning models lack interpretability for clinicians and require substantial GPU resources. This paper presents the first application of BCK-GAN — a bi-directional cascade Gaussian kernel feature-generative adversarial network originally designed for brain tumor detection — to four-class Alzheimer's staging from MRI images. We pair BCK-GAN's feature extractor with a fine-tuned AlexNet classifier trained with Adam, and integrate Grad-CAM to produce visual explanations for each prediction. On a held-out set of 1,000 MRI scans, the pipeline achieved 97.5\% accuracy, outperforming ResNet-50 and DenseNet-121 under identical conditions. While the dataset and protocol remain limited, the results indicate that BCK-GAN's features are discriminative for Alzheimer's staging, and that the approach can deliver accuracy with interpretability on modest hardware.

09:30
09:45

Paper ID: 125

Α Comparative Study of Classical and Generalized Penalized Techniques and Machine Learning Techniques in Linear Modeling

Dimitris Raisis; Alex Karagrigoriou

Presentation: Online

Abstract: Prediction of continuous response variables is an important topic in modern statistical analysis and machine learning. In addition to classical linear models, penalized regression methods such as Ridge, Lasso, and Elastic Net have been developed to address problems including multicollinearity and dimensionality reduction. At the same time, regression tree methods provide a flexible non-parametric approach capable of modeling nonlinear relationships among variables, while modern techniques such as Random Forests combine multiple trees in order to improve predictive accuracy. This work studies and compares the predictive performance of the aforementioned methods, as well as two proposed hybrid approaches that combine variable selection through Elastic Net with regression trees. More specifically, generalized penalty functions for the tree pruning process are investigated in order to examine whether these generalizations can improve the predictive performance of the resulting models. The methods are evaluated using real-world datasets and appropriate performance measures, with the aim of highlighting the advantages and limitations of each approach.

09:45
10:00

Paper ID: 128

Direct Preference Optimization for Decision Making

Antonio Segura; Ziwei Shu; Ramón Carrasco

Presentation: Online

Abstract: Direct Preference Optimization (DPO) was originally introduced as a method for aligning large language models with human preferences expressed over alternative outputs. Many decision-support processes have a structurally similar form: a model proposes candidate decisions, while a human decision maker can express comparative preferences among them. In this work, we study how DPO can be instantiated in probabilistic prescriptive decision-making systems with a discrete set of candidate alternatives. We frame this adaptation using Raiffa’s prescriptive decision theory, specify the requirements needed to apply DPO outside language generation, and describe how conventional scoring models can be converted into probabilistic policies over feasible decision alternatives. We illustrate the framework with a fully synthetic mortgage pre-approval toy example in which pairwise expert preferences adapt the relative probabilities of three discrete credit offers while the original model is retained as a reference policy. The contribution is conceptual and methodological rather than a new credit scoring model or an empirical underwriting study. Extending the formulation to continuous decision spaces is left for future work.

10:00
10:15

Paper ID: 129

Improving Brain Tumor Segmentation in Low-Resource MRI via Diagnostically Relevant Contrast Enhancement

Dominik Toman; Miroslav Laco; Wanda Benesova

Presentation: Online

Abstract: Closing the performance gap between high- and low-resource MRI populations usually requires either retraining the segmentation network or applying a generic, content-agnostic correction filter – here we show a third approach is possible. Deep learning models for brain tumor segmentation (e.g. nnU-Net) trained on high-resource MRI data such as BraTS-GLI degrade measurably on low-resource and lower-quality MRI such as the BraTS-Africa (BraTS-SSA) cohort, due to differences in scanner field strength, coil hardware, and acquisition protocol. Retraining on the target domain is usually infeasible given the scarcity of expert-annotated local data. We propose a test-time, zero-retraining enhancement framework that improves the input image quality to match the intensity statistics (contrast) in which the original segmentation network was trained, leaving the network itself frozen. Rather than a generative correction model, we quantitatively characterize the related BraTS-GLI/BraTS-SSA domain quality gap and design a deterministic, closed-form correction with zero gradient-trained, mechanismspecific parameters, applied only to the modalities and tumor subclasses for which the measured gap is reasonable and stable. The mechanism is calibrated on a heldin population via a leak-free split-half protocol, validated via a pre-registered test, and confirmed on a fully independent test cohort. On this held-out set, the mechanism shows a significant, patient-level Whole Tumor Dice increase of +0.73% (Wilcoxon 𝑝 = 0.0077, surviving FDR correction), with Tumor Core and Enhancing Tumor improving. An independent second run on a separately-trained network pair reproduces the same pattern (+2.01% WT), confirming the effect is not an artifact of one training instance.

10:15
10:30

Paper ID: 132

Explainable Machine Learning for Adaptive Identification of Felder–Silverman Learning Profiles in Programming Education

William Ruiz Martinez; Eugenio Cesar Velázquez Santana; Luis Jose Muñiz Rascado; Wendy Stephanie Martínez Almendares

Presentation: Online

Abstract: Artificial intelligence (AI) and machine learning (ML) are increasingly used to support adaptive learning and personalized educational interventions. However, many studies emphasize predictive accuracy without jointly examining class imbalance, calibration, robustness, explainability, subgroup fairness, and expert validation. This study presents and evaluates a reproducible ML pipeline for identifying four Felder–Silverman learning-profile dimensions: Active/Reflective, Sensing/Intuitive, Visual/Verbal, and Sequential/Global from adaptive-learning constructs and item-level indicators. The final dataset contained 158 students and 33 eligible predictors after an explicit anti-leakage audit. Three feature representations and three classifiers were evaluated using five-fold stratified cross-validation and out-of-fold (OOF) predictions. Model selection used Macro-F1 as the primary metric, complemented by minority-class F1, Matthew’s correlation coefficient (MCC), balanced accuracy, ROC-AUC, calibration, bootstrap confidence intervals, ablation, seed sensitivity, subgroup fairness, and SHAP explanations. The four selected classifiers improved Macro-F1 over majority-class baselines by 0.071–0.106. The strongest gain was obtained for the Sequential/Global profile (Macro-F1=0.563; Δ=+0.106). Expert validation of the adaptive decisions involved seven experts, 30 cases, and five criteria, yielding S-CVI/Ave=0.985. In contrast, all four continuous-score regression targets produced negative R² values relative to the mean baseline. Therefore, the evidence supports predictive classification and expert-validated personalization, but not yet causal educational effectiveness, temporal stability, or external generalization. The study contributes a reproducible and explainable evaluation framework for adaptive programming education.

10:30
10:45

Paper ID: 145

Homophily and the Emergence of Community Structure: An Agent-Based Study of Intercommunity Brokerage

FABIANE CARVALHO; Ivan Bergier; Jayme Barbedo

Abstract: Collective action in institutional networks depends on both cohesion within groups and connections across group boundaries. We investigate how homophilic preferences shape community structure and intercommunity brokerage through an agent-based model. The model represents 1,000 agents initially distributed among five behavioral groups, whose tie formation is governed by a homophily parameter $H$, external-group benefits, and stochastic variation. We vary $H$ from 0 to 5 and perform 20 simulations for each value, measuring behavioral and structural mixing, modularity, community correspondence (ARI and NMI), and brokerage. Results reveal a pronounced structural transition. Under low homophily, intergroup mixing is high and behavioral groups do not correspond to emergent network communities. As homophily increases, behavioral groups become structurally expressed, modularity and community correspondence rise sharply, while boundary nodes decrease and brokerage becomes concentrated among fewer connectors. Strong homophily ultimately produces highly segregated communities with negligible intercommunity connectivity. These findings highlight how micro-level preferences can reorganize meso-level network structure and brokerage.

10:45
11:00

Paper ID: 141

SENTINEL: Dynamic Replanning in Autonomous UAVs Under Pop-up Threats: A Hybrid RRT* and ACO Approach

Selma Akkok; Ozan Duru; Ipek Bengisu Gulper; Ozgur Koray Sahingoz

Presentation: Online

Abstract: This paper presents SENTINEL, a real-time UAV trajectory planning and replanning framework for fixed-wing unmanned aerial vehicles operating in dynamic threat environments. The system combines a global route optimizer based on elitist Ant Colony Optimization (ACO) with a local replanning module built on Rapidly-exploring Random Tree Star (RRT*). A Python planning back-end and a Unity 3D simulation front-end communicate over Unix Domain Sockets using a 4-byte lengthprefixed framing protocol, keeping inter-process communication latency negligible relative to planning time. Global route optimization is evaluated on the Berlin52 benchmark (a standard 52-city TSP instance projected onto a procedural 3D terrain); the elitist ACO with 2-Opt postprocessing achieved an 8.4% route cost reduction over 25 iterations in an obstacle-rich scenario without any algorithm-level changes for the threatzone case. During simulated flight, six emergency replanning events were handled at an average latency of 6.7 ms and a worst case of 15.9 ms, well within the 100 ms real-time budget. The mission completed with zero collisions, a 100% obstacle avoidance rate, and a trajectory tracking RMSE of 13.79 m across an 850 × 850 m map.

11:00
11:15

Paper ID: 100

Onion Seed Quality Prediction Using XGBoost Tuned by an Adaptive Sinh Cosh Optimizer

Veljko Spasic; Bojan Zimonjic; Tea Dogandzic; Andjela Jovanovic; Komlen Lalovic; Milos Grubjesic; Nikola Dragovic; Samson Omasirichukwu Offorjindu; Nebojsa Bacanin; Milos Antonijevic

Presentation: Online

Abstract: This research focuses on predicting onion seed quality from measurable attributes of the seed lot. Since the available public onion seed quality dataset was found to be too simple for a meaningful comparison of optimization methods, a synthetic dataset was generated while preserving the original sample count, feature structure, class proportions, and feature means, but introducing stronger overlap between the quality classes. Classification is performed with XGBoost, while simultaneous feature selection and hyperparameter tuning are carried out by the proposed Adaptive Sinh Cosh Optimizer (ASCHO). Matthews correlation coefficient (MCC) was used as the optimization objective. XG-ASCHO was compared with the original SCHO and four additional metaheuristics over 30 independent runs. The proposed method achieved the best individual MCC of 0.5255, as well as the highest accuracy (0.617), macro-F1 (0.519), and weighted-F1 (0.484) among the best models obtained by the compared optimizers. SHAP analysis of the best XG-ASCHO model showed that germination rate is the dominant feature influencing the predictions, while the remaining attributes have considerably smaller contributions. The obtained results demonstrate the potential of the proposed optimization and explainability framework for onion seed quality assessment in precision agriculture.

Technical Session - 6 Online Time: 09:00 - 11:15. Location: Virtual Room - 2

Join Session

Session Chairs: Dr. Volkan Abur

09:00
09:15

Paper ID: 146

Mobile Technologies and IoT Devices in Smart Dementia Care

Aslı Aslantaş; Buse Tatlıses; Özgür Koray Şahingöz

Presentation: Online

Abstract: Mild Cognitive Impairment, dementia, and Alzheimer’s disease are diseases which involve cognitive and functional decline and which could impair someone’s independence and place great strains on health systems. Mobile technology, wearable devices, IoT, and AAL have recently been receiving increasing attention in relation to their uses in the management and monitoring of MCI, dementia, and Alzheimer’s disease. In this paper, the applications of such technologies in various areas, including behavior and activity monitoring, location and missing person detection, fall and emergency management, drug administration and cognition assistance, and remote monitoring, are explored. Also we have included a discussion on the role of AI and machine learning in context awareness, personalization, and edge and cloud computing concepts. From the literature, it is clear that the applications of such technologies will enable continuous monitoring of patients, and they will help with the activities of daily living, assist in identifying risks associated with safety issues, and even allow communication between the patient and caregiver. However, there are a number of issues still remaining as a barrier for the use of these technologies, including the privacy and security of data, ethical issues, acceptability and usability of systems, interoperability, and reliability of the systems. These concepts discussed above give a comprehensive review of the application of mobile technologies, wearables, IoT technologies, and smart home technologies to dementia and Alzheimer’s disease patients. This paper aims to provide a thorough background of the current research on these technologies and get a better understanding of how emerging technologies can help in monitoring and assisting dementia and Alzheimer’s patients.

09:15
09:30

Paper ID: 147

A Multi-Criteria Decision Model for Computational Offloading in Hybrid Cloud-Edge Systems Accounting for Container Cold-Start Cost

Oleksandr Anatskyi; Andriy Dudnik; Andrii Fesenko; Dmytro Zhuravlov

Presentation: Online

Abstract: Conventional task offloading in hybrid cloud-edge systems is typically optimized for network latency, computation time, and energy consumption, under the implicit assumption that the containerized execution environment is pre-deployed at the target node. In practice, this assumption is often invalid, leading to cold-start latency that existing offloading models do not address. This study describes a multi-criteria offloading decision model that explicitly includes the cold-start cost – a function of the local image-registry state and channel bandwidth – into the latency objective. The hybrid system is formalized as a weighted directed graph G = (V, E, W), and the offloading problem is formulated as a three-vector integer optimization over latency, energy consumption, and load balancing, subject to six constraints. The problem's NP-hardness is proven by reduction from Partition. An adapted NSGA-III algorithm (NSGA-III-CA) with a cache-aware mutation operator is developed, and a formal guarantee is provided that bounds the expected increment in cold-start cost relative to standard mutation. Simulation studies across three problem scales and four experiment series demonstrate that the proposed method removes systematic scheduling bias by construction and statistically significantly reduces the deadline violation rate compared to four baseline algorithms, with a moderate trade-off in energy efficiency and an 8–24% increase in runtime relative to standard NSGA-III.

09:30
09:45

Paper ID: 148

Measuring Error-Sensitive Human Oversight in AI-Assisted Decisions: A Cross-Domain Stress Test with a Hiring Replication

Amirreza Eshraghi; Jelena Vujicic; Bahareh Kudarzi

Presentation: Online

Abstract: Regulatory frameworks increasingly require meaningful human oversight of high-risk artificial intelligence (AI) systems used in employment and hiring, including the European Union Artificial Intelligence Act (Article 14) and New York City Local Law 144. These frameworks do not, however, specify how to determine whether oversight has actually been exercised. We isolate and test one observable behavioral component of oversight: error-sensitive intervention, defined as correction of AI recommendations that are independently classified as erroneous. We evaluate this measure on two independent public datasets of real human-AI decisions in which the correctness of each AI recommendation can be established: a criminal-judgment task (260 participants) and a hiring task (349 participants with at least one eligible trial). Acceptance of an erroneous recommendation is treated as an uncorrected AI error, not as proof of mindless deference. We find that uncorrected AI errors are common (22.7% and 54.2% of participants accepted at least one erroneous recommendation); that in the hiring task participants exhibited calibrated reliance, following score-consistent AI on 90.3% of trials but score-discordant AI on only 35.5%; and that error-sensitive intervention predicts decision quality on disjoint trial sets (split-partition mean rho = 0.28), establishing information beyond same-trial mechanical coupling. The measure carries real but incomplete information about decision quality and is not a universal compliance score. We argue oversight requirements should specify graded, independently-adjudicated behavioral measures interpreted alongside behavior on correct recommendations and outcome quality.

09:45
10:00

Paper ID: 150

Neurocontrollers in Autonomous Robotics supported by Feeling AI: Training Methods and Control Schemes

Anatolii Kargin; Tetyana Petrenko; Olena Piatykop; Olha Pronina

Presentation: Online

Abstract: Autonomous Robotic Systems (ARSs) supported by hybrid AI are in demand today. At low levels of ARSs control system, many different types of feedforward and feedback controllers are used. Integrating these controllers with hybrid AI offers advantages for ARS. The purpose of this paper is to demonstrate how, by synthesizing low-latency reactive response of Neuro Controllers (NCs) at the low-level of the ARS, event-driven spiking architectures, and drive-emotion cognitive loops, Feeling Artificial Intelligence (FAI) enables real-time, low-level neural control on energy-restricted embedded hardware without the need for massive retraining using cloud computing. The operational and learning schemes of conventional multilayer Artificial Neural Networks (ANNs) suitable for integrating with the FAI components are discussed. The sequential and parallel learning schemes for feedforward and feedback NCs are presented; the different options for their implementation in ARS are illustrated. It is shown that, under FAI supervision, the training procedure of the conventional ANN used as ARS’s NC may be similar to that of a Spiking Neural Network (SNN). The experiments with the feedback NC, including those in the drive-emotion cognitive loop of FAI that performs the tracking task, are discussed.

10:00
10:15

Paper ID: 110

Surrogate-Assisted Particle Swarm Optimization of Cascaded PI Controllers for a Two-Stage Solid-State Transformer

Turgut Mert Aksoy; Enver Ulusoy; Ceyda Cogal; Hidir Duzkaya

Presentation: Online

Abstract: Solid-state transformers (SSTs) intended for smart grid-oriented low-voltage DC (LVDC) distribution rely on cascaded proportional–integral (PI) controllers spanning a power-factor-correction (PFC) front end and an isolated phase-shifted full-bridge (PSFB) output stage, giving rise to three interacting loops and six coupled gains that cannot be tuned independently through conventional loop-by-loop design. This paper proposes a surrogate-assisted framework for jointly tuning these six PI gains on a 1 kW two-stage SST prototype (360 V DC-link, 48 V / ~23 A output). A scalar cost function over a six-dimensional gain space combining power factor, total harmonic distortion (THD), output-voltage ripple, and reference-tracking error is minimized using an 800-point Latin hypercube sampling (LHS) campaign in PLECS, from which a multilayer perceptron (MLP) surrogate is trained (R² ≈ 0.90 for the cost function, R² ≈ 0.94 for THD) and used to guide particle swarm optimization (PSO); a full PLECS re-simulation confirms the resulting design. The PLECS-verified PSO design (J = 0.294, PF = 0.996, THD = 5.54%, ripple = 0.29 V) substantially outperforms the hand-tuned baseline (J = 0.436, ripple = 1.08 V), mainly through reduced output ripple, but does not surpass the best individual LHS sample (J = 0.285). The PSO solution saturates at the search-box bounds on three of the six gain axes. These results show that surrogate-assisted PSO is a computationally practical alternative to simulator-in-the-loop search for multi-loop SST controller tuning, and that a sufficiently dense LHS sweep alone can locate a near-optimal design within the explored search region.

10:15
10:30

Paper ID: 163

Controlled Comparison of Filter, Wrapper, and Embedded Feature Selection Approaches for Network Intrusion Detection

Elif Yıldırım; İsa Avcı

Presentation: Online

Abstract: The growth of network infrastructures and the increasing number of connected devices have intensified the complexity of cyber threats and increased the demand for effective intrusion detection systems (IDS). Although machine learning-based approaches are widely used in this domain, the high-dimensional feature structure of modern network datasets poses challenges in terms of model complexity, training time, and computational cost, making feature selection a critical research problem. In this study, the effects of filter-based Mutual Information and Chi-Square, wrapper-based RFE, and embedded Random Forest Feature Importance methods on intrusion detection performance were comparatively analyzed using the UNSW-NB15 dataset. To evaluate the independent effect of each method, the same Logistic Regression classifier was used in all experiments, and feature subsets of 𝑘 = 10, 20, 30, 40, and 50 were generated. The results show that appropriate feature selection has a significant effect on IDS performance; the Mutual Information method achieved the best results with 50 features, reaching 81.2% accuracy, 75.2% precision, 98.4% recall, and an F1-score of 85.3%, along with a ROC-AUC of 0.955 and an Average Precision of 0.967, demonstrating strong discriminative capability. The findings indicate that, without requiring highly complex models, effective feature selection strategies can enable the development of lower-cost and more practical IDS solutions

10:30
10:45

Paper ID: 102

Personalized Delivery of Instructional Materials

Nikolai Scerbakov; Vadim Pak; Vladimir Muliukha

Presentation: Online

Abstract: Traditional Learning Management Systems (LMS) often adopt a rigid, one‑size‑fits‑all course structure that fails to address the diverse needs of individual learners, leading to disengagement, confusion, and suboptimal outcomes. This paper proposes a personalized tutoring framework that overcomes these limitations by delivering adaptive, learner‑centered educational content. The framework leverages the HM‑Data model to structure courseware into modular, addressable units called S‑collections - self‑contained mini‑courses that encapsulate instructional materials, assignments, and assessments within an internal navigational structure. A push‑based distribution system delivers these personalized containers to students via their preferred communication channels, including email, social network bots, and web push notifications. A key contribution is the extension of delivery beyond server‑centric browsing to offline‑capable formats. S‑collections can be automatically compiled into EPUB files - an open, standards‑based e‑book format with broad cross‑platform support - within seconds. These packages can embed personalized hints, examples, and supplementary readings, and can be opened in any standard EPUB reader, enabling students to study independently of internet connectivity. This approach enhances accessibility, reduces server load, and supports features such as annotation and responsive layout adaptation.

10:45
11:00

Paper ID: 166

Deep Learning for Acute Lymphoblastic Leukemia Detection: A Multi-Architecture Comparison of CNN and Transformer Models on Microscopic Blood Smear Images

Zuhaa Nisar; Muhammad Atif Saeed; Akhtar Jamil; Alaa Ali Hameed

Presentation: Online

Abstract: Acute Lymphoblastic Leukemia (ALL) is a fast- growing, blood and bone marrow cancer, and the most prevalent cancer in children globally. The early and accurate classification of ALL subtypes is important to effectively manage patients’ care; manual analysis of peripheral blood smear images is time- consuming, subjective, and cannot always capture subtle mor- phological differences. In this study, a deep learning framework based on the transfer learning approach for automated early detection and multi-class classification of ALL is presented, employing the state-of-the-art transfer learning architectures. A publicly available dataset consisting of 3,256 microscopic images was used and categorized into four classes: Benign, Early Pre- B, Pre-B, and Pro-B. Four modern architectures, EfficientNet- B0, EfficientNet-B3, ConvNeXt-Tiny, and Custom Vision Trans- former, were evaluated using a stratified 70/15/15 split and a two-stage fine-tuning strategy with class-weighted cross-entropy loss. A detailed ablation study has been performed to evaluate the impact of augmentation strength, class weighting, and optimizers (Adam, Adamax, SGD with momentum). Experimental results indicate that ConvNeXt-Tiny outperforms other models in terms of test accuracy at 99.39%. On the other hand, EfficientNet- B0 offers the best balance of accuracy (98.77%), computational efficiency, and deployment suitability. The best performing CNNs targeted clinically relevant regions, as was confirmed by Grad- CAM visualizations, which included chromatin patterns in the nucleus as well as nuclear-to-cytoplasmic ratios. The proposed framework achieves the accurate, interpretable, and compu- tationally efficient automated classification of ALL subtypes, significantly outperforming the baseline CNN methods.

Technical Session - 7 Online Time: 11:15 - 13:30. Location: Virtual Room -1

Join Session

Session Chairs: Dr. Ruslan Isaev, Dr. Ayşe Su

11:15
11:30

Paper ID: 167

Toward Grounded Qur’anic Question Answering: An Agentic Retrieval Pipeline and Empirical Audit of Embeddings, Abstention, and Citation Guardrails

Laiba Noor; Rooma Maryam; Muhammad Atif Saeed; Akhtar Jamil; Alaa Ali Hameed

Presentation: Online

Abstract: Large language models can answer questions about the Qur'an fluently, but fluency is not evidence of textual support. This paper presents and audits an agentic retrieval-augmented generation (RAG) pipeline for Qur'anic verses and Tafseer. The system combines Arabic semantic retrieval, verse-level provenance, bounded retrieve--assess--refine iterations, citation-aware generation, and a graph-inspired structural check. The experiments show a clear but limited result. On 174 AyaTEC questions, base GATE-AraBERT-v1 with a verses-only index achieved HitRate@20 of 0.6724 and MRR of 0.3626; on 169 QRCD questions, it achieved MRR of 0.3015. The tested verse--Tafseer fine-tuning variants were weaker on question-to-verse retrieval, and a mixed verse--Tafseer index reduced recall. Correcting the HNSW search setting from $ef=10$ to $ef=256$ improved recorded Recall@10 by 18.9\%. The safety experiments did not justify a zero-hallucination claim: similarity-based abstention over-refused 88.3\% of answerable controls at one high-rejection threshold, while the structural guardrail matched the majority baseline on a nine-answer audit. The contribution is an evidence-led account of a promising retrieval core, its failure modes, and the repairs required before dependable grounded QA can be claimed.

11:30
11:45

Paper ID: 17

Evaluating the Psychosocial Impact of AI Social Robots on Children with High Support Needs

Cheng Kian TAN

Presentation: Online

Abstract: The rapid ageing of populations has increased interest in socially assistive robots (SARs) as potential tools to support older adults and caregivers in eldercare settings. While prior studies have focused largely on care recipients, less is known about how caregivers perceive the long-term integration of SARs into everyday care practice. This study examined professional caregivers’ perspectives on the use of LOVOT, a socially assistive companion robot, as an adjunct psychosocial support tool in eldercare. A qualitative longitudinal design was adopted involving semi-structured interviews with seven pro-fessional caregivers across community sheltered homes and senior daycare centres in Singapore. Interviews were conducted after an initial 12-session intervention involving 105 older participants and again during a post-study follow-up phase. Thematic analysis revealed that caregivers consistently viewed LOVOT as a complementary rather than replacement form of care. LOVOT was perceived to support emotional regulation, activity participation, social engagement, and short periods of operational respite for care staff. Longitudinal findings suggested that engagement with LOVOT extended beyond nov-elty effects and became increasingly integrated into routine care practices over time. However, caregivers also highlighted implementation challenges, including training needs, battery limitations, workflow integration, supervision requirements, and the need for greater personalization. Overall, the findings suggest that LOVOT may function as a useful psychosocial support tool when embedded within a supportive care ecology and integrated into staff-led caregiving practices.

11:45
12:00

Paper ID: 23

Enhancing IoT Security: A Hybrid Deep Learning, En-semble, and Transfer Learning Framework for IDS

Yahya Alhaj maz; Mahmoud Alahmad; Muaz Zakaria Mohamad Hajmoez

Presentation: Online

Abstract: The Internet of Things (IoT) represents a burgeoning concept linking physical entities to the digital realm via the Internet. However, the rapid proliferation of IoT devices brings forth notable security and privacy concerns, given the inadequacy of protective measures for many of these devices. Exploiting vulnerabilities in IoT devices and networks, malicious actors can execute various forms of attacks. This study conducts a comprehensive examination of attacks targeting IoT networks and perception layers. It introduces a classification framework for intrusion detection systems within IoT networks, focusing on ensemble methods, deep learning techniques, and transfer learning. Furthermore, it conducts a comparative analysis of this framework against existing surveys, evaluating aspects such as objectives, scope, and methodology. Lastly, the survey outlines research gaps and unresolved issues to guide future endeavors in this domain.

12:00
12:15

Paper ID: 27

A holistic theoretical framework for non-player character personality and its social network model

Muhtar Çağkan Uludağlı; Kaya Oğuz

Presentation: Online

Abstract: In modern computer games, non-player characters (NPCs) often lack unique personalities and engaging, compelling behaviors, which can significantly diminish the player's immersive experience. To effectively address this prevalent issue, the current study proposes the development of a novel theoretical model for NPC personality alongside an innovative adaptive NPC social network model. Specifically, the proposed personality model aims to significantly enhance NPCs' character traits by incorporating various sub-models that address the personality, actions, memory, learning, communication, and health facets of the overall NPC model comprehensively. Additionally, this paper provides a comprehensive review of the related literature, examining general NPC personality models and explaining the foundational concepts that support them. Furthermore, a proposed adaptive NPC network model is strongly recommended to more accurately simulate real-life social interactions and connections between individuals. The objective of this innovative approach is to promote more realistic and dynamic relationships among NPCs, which could potentially enhance the overall realism and engagement of the gaming environment, thus improving the player's experience.

12:15
12:30

Paper ID: 34

An XAI-Ready Rule-Based Clinical Data Modeling Framework for Early Referral-Oriented Assessment of Prodromal and Non-Bullous Pemphigoid

Ebru Erdem; Merve Hatun Erkayman

Presentation: Online

Abstract: Bullous Pemphigoid (BP) is an autoimmune bullous disease that may initially present with nonspecific clinical findings before the emergence of tense bullae. This phase can complicate early differentiation of the disease from chronic urticaria and eczema-like presentations. Although the dermatology literature has increasingly emphasized the diagnostic value of clinical history, lesion duration, biopsy, Direct Immunofluorescence (DIF) examination, and serological markers, there remains a need for studies that translate these heterogeneous clinical variables into a structured, explainable data model suitable for future clinical decision support systems. This study proposes an explainable clinical data modeling framework for the early differential assessment of prodromal/non-bullous BP, a presentation frequently mistaken for urticaria-like clinical findings. Serological autoantibody markers—anti-BP180 (anti–Bullous Pemphigoid Antigen 180) and anti-BP230 (anti–Bullous Pemphigoid Antigen 230)—are considered alongside Immunoglobulin G (IgG) and Complement Component 3 (C3) deposits detected by DIF. Together, these markers are instead positioned as structured, complementary data fields for post-referral diagnostic confirmation and future model validation. The framework integrates rule-based scoring, a weighted decision matrix, and a maturity roadmap for future real-world validation studies. The functionality of the proposed structure is illustrated through hypothetical clinical scenarios. The original contribution of this study lies in transforming clinical diagnostic reasoning into a structured, explainable, and validation-ready data model. Future work should validate the framework using retrospective or prospective patient datasets obtained under institutional ethics committee approval.

12:30
12:45

Paper ID: 114

Every Token Counts: Cost-Aware Multi-Agent CTF Solver with MCP

Ahmet Bucko; Kastriot Fetahaj; Blerim Rexha

Presentation: Online

Abstract: Large language models (LLMs) are rapidly evolving and becoming more deeply embedded in everyday life, reshaping many previously settled practices along the way. One such practice is the Capture-the-Flag (CTF) challenge-solving competition. Where such competitions were once fully manual, rewarding the most skilled and experienced human players, the field has since shifted toward rewarding whoever builds the better agents and tools. In this paper, we present our MCP CTF solver, a system built on the Model Context Protocol (MCP) architecture that connects multiple CTFd- and web-based clients to specialized agents and tools within a single solver, with agent-to-platform communication already handled natively through MCP.

12:45
13:00

Paper ID: 38

A Conceptual Framework for Adaptive and Explainable Smart Home Energy Management

Kalin Petrov; Todor Rachovski; Emil Hadzhikolev; Stanka Hadzhikoleva; Kostadin Yotov

Presentation: Online

Abstract: Rising electricity prices and increasing market complexity necessitate more effective management of household electricity consumption. Home Energy Management Systems (HEMS) provide a technological foundation for such management; however, their practical effectiveness critically depends on their analytical capabilities. This study presents a comparative analysis of four representative HEMS platforms – Google Nest, Sense, Smappee, and Home Assistant – evaluated against five key criteria: presence of an AI-based forecasting component, model adaptability, explainability of recom-mendations with quantified financial impact, hardware independence, and compatibility with local tariff structures. The results show that none of the analyzed systems simultaneously satisfies all criteria, with each platform covering only a subset of functionalities. Based on these findings, a concep-tual model, EnergyForecast, is proposed. The model integrates four core components: a hybrid forecasting engine based on XGBoost and LSTM, an adaptive mechanism for concept drift detection and automatic retraining, an explainable recommendation module with quantified financial impact, and an interactive simulation module for scenario-based planning. The proposed approach is designed to be hardware-independent and adaptable to local en-ergy market conditions and outlines a direction for the development of next-generation HEMS systems integrating forecasting, explainability, and market adaptation.

13:00
13:15

Paper ID: 4

Sovereign AI for Science: Linking Infrastructure Control to Scientific Reproducibility

Otmane Azeroual; Said Oualibouch

Presentation: Online

Abstract: Artificial Intelligence (AI) has become a core instrument in scientific discovery, yet its growing reliance on externally governed systems introduces structural risks to reproducibility, transparency, and long-term scientific reliabil-ity. Existing AI governance approaches primarily address ethical and regulatory concerns but insufficiently account for the infrastructural conditions required for reproducible AI-enabled science. This paper introduces Sovereign AI for Science as a framework that links verifiable control over data, models, and computational infrastructure to scientific reproducibility and auditability. We argue that AI systems must be treated as epistemic artefacts whose lifecycle must remain accessible and controllable over time. We operationalize this concept through a three-pillar framework—data control, model control, and infrastructure control—and derive a layered reference architecture for reproducible and auditable AI-enabled research. Furthermore, we analyse how non-sovereign AI setups lead to reproducibility failures, including temporal instability, restricted methodological scrutiny, and loss of longitudinal comparability. The paper concludes with implications for research policy and outlines a research agenda for embedding sovereign AI principles into scientific infrastructures.

13:15
13:30

Paper ID: 168

An Experimental Framework for Life Expectancy Prediction from Single-Country Life-Factor Data

Ruslan Isaev; Asina Bopushova

Abstract: Life-expectancy prediction from person-level risk factors is usually studied on a small cohort from one country, which limits data volume and transferability. We propose an experimental framework that combines any person-wise life-factor dataset with public life tables to predict life expectancy for the reference country and for every country with a life table. A Kyrgyz cohort of 95 participants, collected with a cyber-physical system, is labelled with the remaining life expectancy $e(x)$ of the 2023 Kyrgyz life table. It is modelled with a mean baseline, Ridge, Random Forest (RF) and Gradient Boosting (GB), and re-labelled with the tables of 192 countries. A built-in leakage audit shows that the cross-validated R$^2$ of 0.96 comes from age (97-98\% of SHAP), and that RF and GB lose their skill without age and its proxies. In country-level forecasts, naive persistence was hardest to beat after the COVID-19 rebound (MAE 0.69 years in 2023). We report prediction status and feature importance for all 192 countries.

Technical Session - 8 Online Time: 11:15 - 13:30. Location: Virtual Room - 2

Join Session

Session Chairs: Dr. Habip Koçak

11:15
11:30

Paper ID: 41

From Reporting to Investment: A Non-Compensatory Framework for SME Green-Transition Readiness

Alper Ozpinar; Bugra Kara; Omer Enging Tuzun; Burcu Kasimoglu

Presentation: Online

Abstract: Small and medium-sized enterprises (SMEs) face a gap between answering sustainability-information requests and preparing a controlled, financeable transition action. This conceptual position paper defines SME Green Transition Readiness (SME-GTR) as a non-compensatory, versioned decision profile. It separates reporting, greenhouse-gas accounting, and financial-assessment capabilities from action-specific headroom, secured-financing coverage, financing feasibility, and process integrity. Enduring cases are linked to immutable versions; every evaluation retains a disposition and package snapshot. Four temporally identified empirical propositions are distinguished from one falsifiable artifact design proposition so that authorization does not determine the outcomes. A synthetic heat-conversion case demonstrates reproducible coverage, liquidity, leverage, and collateral calculations. The information architecture resolves evidence, calculation, scenario, and commitment flows. SME-GTR is unvalidated; pre-registered content, reliability, feasibility, and confirmatory studies are required before empirical use.

11:30
11:45

Paper ID: 103

Predicting Decision Analysis Time in Image-Guided Radiotherapy Using Machine Learning

Shahad Alrushud; Hamed Alshamri; Ali Alghamdi

Presentation: Online

Abstract: Decision analysis time in image-guided radiotherapy (IGRT), defined as the interval between image acquisition and application of the corresponding setup correction, directly affects treatment efficiency, room throughput, and patient comfort. However, it has received little attention in predictive modeling compared with contouring, treatment planning, and quality assurance. This study retrospectively analyzed 899 IGRT fractions from Prince Sultan Military Medical City, Riyadh, incorporating clinical, treatment, and workflow variables to predict decision analysis time. Random Forest, Gradient Boosting, XGBoost, and a TensorFlow neural network were evaluated, with Bayesian hyperparameter optimization (HyperOpt, TPE) applied to the tree-based models. The mean decision analysis time was 5.12 minutes (SD = 4.65). Gradient Boosting achieved the best regression performance (MSE = 6.07, R² = 0.49). For classification into short (≤5 min) and long (>5 min) categories, baseline accuracies ranged from 0.78 to 0.82, increasing to 91.1% for the optimized Random Forest and XGBoost classifiers, while the neural network achieved 89% accuracy. The optimized models generalized to 13 unseen cases, achieving up to 84% accuracy. Session repetition and average isocenter shift were consistently the strongest predictors. These results demonstrate that ensemble machine learning can reliably model decision analysis time in IGRT, providing a practical approach to improving scheduling and workflow efficiency in radiotherapy departments.

11:45
12:00

Paper ID: 117

Mining Multi-Source Agent Traces for Orchestration Skeletons: Weak Supervision and Shortcut Diagnostics

Joohyun Lee; Wen-Syan Li

Presentation: Online

Abstract: Public execution traces of large language model (LLM) agents record tool calls, code repairs, retrieval decisions, and multi-agent handoffs, but they rarely label the pre-execution coordination structure that an orchestration runtime should use. We study whether heterogeneous traces can be normalized into auditable weak supervision over candidate orchestration skeletons. A skeleton specifies roles, topology, synchronization, permissions, context scope, and escalation before concrete prompts, tools, and provider adapters are bound. Our pipeline canonicalizes multi-source traces, creates rule-checked counterfactuals and rare positives, composes replay-stabilized training mixtures, and adapts a compact Gemma model with LoRA to reconstruct candidate labels and pairwise weak preferences. On a 30,674-example weighted corpus, stress-aware replay raises cross-source interaction candidate exact match from 0.237 to 0.445 and tracked LTS-II axis accuracy from 0.769 to 0.874 while maintaining 0.885 in-domain candidate exact match. Shortcut diagnostics sharply bound this result: marker-stripped CSI-Stress candidate exact falls to 0.053 and hard-pairwise accuracy is near chance. We additionally report a 36-run exploratory 𝜏-bench execution pilot using independently authored native outcomes. A task-aligned soft workflow prompt achieves 6/12 successes versus 8/12 for both a minimal prompt and a reverse-order control; the observed gap is confined to the second trial and is strongly entangled with user-simulator termination behavior. We therefore treat the pilot as a negative boundary result, not evidence of runtime superiority. Overall, the learned model is best interpreted as a descriptor-conditioned compatibility prior over mined candidates whose external deployment value remains unestablished.

12:00
12:15

Paper ID: 55

An Industry 5.0-Ready Hybrid Digital Twin Architecture for MES

Ilgin Safak; Jari Partanen; Niko Candelin; Majid Haghparast; Tommi Mikkonen; Tapio Frantti; Fatih Alagöz

Presentation: Online

Abstract: The increasing complexity of Industry 4.0 manufacturing systems and the adoption of cyber-physical technologies have exposed limitations of conventional Manufacturing Execution Systems (MES) in supporting scalable, secure, and intelligent manufacturing operations. Although Digital Twins (DTs) enhance MES by enabling virtual modeling, monitoring, and optimization of physical processes, their integration presents significant challenges in interoperability, scalability, security, and orchestration. This paper proposes an Industry 5.0-ready hybrid MES–DT reference architecture that serves as a design blueprint for integrating MES, DTs, edge–cloud computing, AI-based multi-agent systems, and quantum computing (QC) within a unified framework. The architecture incorporates security-by-design, Zero Trust (ZT), service-oriented principles, and post-quantum cryptography (PQC) to address current and emerging cyber threats while supporting distributed and resilient manufacturing environments. Multi-agent systems provide decentralized coordination of manufacturing services, whereas QC is envisioned to accelerate computationally intensive tasks such as production scheduling and optimization. Rather than evaluating a specific implementation, the proposed reference architecture synthesizes established technologies, security mechanisms, and architectural principles into a coherent framework intended to guide the design and future implementation of secure, scalable, and quantum-ready MES–DT ecosystems.

12:15
12:30

Paper ID: 120

Task-Driven Dynamic Agent Aggregation: Joint Reasoning Beyond Routing in Multi-Agent Systems

Hyerim Jeon; Wen-Syan Li

Presentation: Online

Abstract: Multi-agent systems (MAS) built on large language models adapt their structure at the level of communication: static frameworks fix roles and the message graph in advance, while dynamic routing selects which agents to activate for a query. In both, the low-level agents stay separate while a task runs, so a request whose subtasks are coupled by shared global constraints—a monetary budget, an effort budget, an output schema—is distributed across several reasoning contexts and is never available to any of them as a joint decision problem. We argue that the operative question is whether low-level agents should stay separate at execution time or be combined for the task at hand, and we propose Task-Driven Dynamic Agent Aggregation (TDAA). TDAA scores a request with an explicit trigger, selects a subset of the agent pool by similarity ranking with a rule-based coverage check, and composes the selection into a temporary Virtual Aggregated Agent (VAA): a single reasoning context holding a merged representation of their prompts, tools, constraints and objectives. On ConstraintWorld, a ten-application environment with 31 APIs and 100 constraint-bearing tasks, TDAA improves task goal completion over the three message-passing baselines by +0.43 to +0.52 in paired task-level tests, uses 24–49% fewer tokens against two of the three, and reduces global constraint violations to a quarter of the next-best system. A monolithic agent holding every tool is the most accurate system we measure; TDAA reaches 83% of its task completion at 57% of its token cost, and that cost ratio widens as the pool grows from 16 to 40 agents. These results show that whether low-level agents are combined for execution, rather than how they are connected, is what governs performance on requests bound by shared global constraints.

12:30
12:45

Paper ID: 63

Explainable Attentive Relational Graph Convolutional Network for Money Laundering Transaction Topology Classification

Emmanuel Okafor; Hamad Alnaimi; Tariq Alshehri; Taqwa Alhaj

Presentation: Online

Abstract: Money laundering is concealed within financial transaction networks through recurring structural patterns such as structuring, layering, and integration. Identifying which pattern a transaction belongs to, rather than flagging it as suspicious only, allows compliance analysts to act on an alert. This study formulates the task as transaction topology classification: every observed transaction in a heterogeneous account graph is assigned to one of four classes, namely normal, structuring, layering and integration. We propose an Attentive Relational Graph Convolutional Network (AR-GCN) that interleaves relation-specific message passing with attention-based neighbor weighting, so that account representations reflect both the type of transaction connecting two accounts and the relative importance of each neighbor. Experiments on a class-balanced graph derived from the public AMLNet corpus, comprising 3,490 transactions over 4,029 accounts, compare AR-GCN against seven graph neural network architectures and four ensemble baselines under 10-fold stratified cross-validation. AR-GCN attains the highest accuracy (0.9830) while RGCN achieves the best F1-score (0.9750); the relational family as a whole outperforms non-relational graph models by more than forty F1 points, identifying explicit relation modeling as the decisive factor. SHAP attributions aggregated across all twelve models rank source-side transaction amount and temporal behavior as the dominant predictors. Inference completes in under 3 milliseconds per fold, demonstrating low inference latency under the experimental test conditions.

12:45
13:00

Paper ID: 105

Imitation and Offline Reinforcement Learning for Alert Triage on Microsoft GUIDE

Emmanuel Okafor; Ibrahim Alzahrani; Saeed A. Alshahrani

Presentation: Online

Abstract: Security operations centers receive more alerts than analysts can examine, motivating automated triage based on historical analyst decisions. We compare imitation learning and offline reinforcement learning (RL) on the Microsoft GUIDE corpus of real enterprise security operations, using its published benchmark split: 13.7 million analyst-labeled evidence rows, 7.4 million reconstructed training transitions, 3.8 million held-out transitions, and ten seeds across seven agents. A recurrent imitation policy achieves the best detection performance (F1 0.680, ROC-AUC 0.825, MCC 0.465). Three of the seven agents—the two reward-weighted recurrent policies and the Decision Transformer—collapse to predicting every alert as positive, retaining an F1 near 0.538 while their MCC is exactly zero, highlighting the limitations of F1 under class imbalance. A controlled pair sharing one architecture and differing only in whether reward enters the loss attributes this collapse to the reward specification rather than the sequence model itself. Weighted importance sampling fails two consistency diagnostics and is therefore excluded from model selection. Finally, a severity-based containment control recovers 55% of the selected policy’s missed attacks using only predicted severity, with a small MCC decrease. These results provide controlled evidence supporting imitation learning over offline RL on this benchmark and identify practical safeguards for evaluating and deploying learned SOC triage policies.

13:00
13:15

Paper ID: 79

A Leakage-Controlled Transformer Baseline for Non-Technical Loss Prediction via Collection Efficiency

Emmanuel Okafor; Patience E. Aigbe

Presentation: Online

Abstract: Non-technical losses constrain the financial viability of electricity distribution companies in developing economies, and in Nigeria the collection component alone exceeds all technical and commercial causes combined. An established line of work predicts collection efficiency bands from monthly billing records using deep learning. We examine the algebraic structure of the task: collection efficiency is collection amount divided by bill amount, both natural members of an operational feature set, making the target an exact function of two inputs. The dependence persists after the conventional multicollinearity correction, since energy consumption explains bill amount with $R^2 = 0.9949$, so reported performance may be attributable in part to this feature-target dependence rather than solely to learned structure; a deterministic baseline computed from the retained features reaches 0.8497 accuracy, giving a reference point for scores under this configuration. Building on that diagnostic, we evaluate three feature configurations spanning this spectrum, comparing a feature-tokenised Transformer encoder (NTL-TransNet) against stacked LSTM and GRU baselines under five seeds by stratified five-fold cross-validation (450 runs). NTL-TransNet attains 0.7961 weighted F1 on FS-A, 0.7711 on FS-B and 0.4285 on the leakage-controlled FS-C, exceeding both recurrent baselines on all 150 paired folds, by a median 0.130 weighted F1 over the stacked GRU on FS-C. Its core strength is stability: its weighted-F1 standard deviation stays at or below 0.043 across the three configurations, where the recurrent baselines reach 0.164. Optimiser differences are significant under the fuller configurations but not under leakage control, and Shapley attributions and encoder attention agree only weakly.

13:15
13:30

Paper ID: 155

Descriptive Pilots Mislead: An Empirical Study of Input Form in Retrieval-Augmented Enterprise Assistants

Tarmo Koppel; Tarmo Koppel

Abstract: Retrieval-augmented generation has made firm-specific virtual assistants builda-ble by business users without engineering support, but evaluations of such assis-tants typically test descriptive question answering over prose and generalize from there. This study asks whether that generalization holds. Seventy-one business professionals each built a retrieval-augmented assistant from documents describ-ing a firm they had personally observed, then evaluated it across three input forms: descriptive prose indexed for retrieval; tabular and numeric inventory and policy documents added to the same index; and an incoming customer purchase order attached to the query rather than indexed. Each participant rated ten answers per form against ground truth they alone possessed, plus 38 Likert items per model condition. Accuracy declined monotonically across the three forms, from 4.35 to 4.11 to 3.97 out of 5 (repeated-measures ANOVA p = .033; Friedman p = .007), the share of near-perfect assistants halved, and within-participant disper-sion rose by two-thirds. The penalty was unrelated to the volume of numeric con-tent. Critically, descriptive performance predicted numeric performance only weakly (r ≈ .21), whereas the two numeric forms correlated at r = .68, indicating that these input forms exercise distinct capabilities and that descriptive pilots are poor evidence for operational deployment.

Technical Session - 9 Online Time: 13:30 - 15:45. Location: Virtual Room -1

Join Session

Session Chairs: Muhammad Atif Saeed, Dr. Kateryna Vodolaskova

13:30
13:45

Paper ID: 86

Joint Analysis of Fault Probability and Age Estimation for High-voltage Insulator Condition Assessment

Waleed Alshaibani; Ramazan Çağlar; Ibraheem Shayea

Presentation: Online

Abstract: This paper presents a joint condition-assessment approach for high voltage insulators by investigating the relationship between fault probability and insulator degradation. The proposed approach builds on previously developed fault prediction and age estimation frameworks and graphically relates normalized fault probability to two degradation representations based on texture–structure and color–structure characteristics. The results showed that fault probability and degradation severity do not exhibit a simple one to one relationship. Especially low fault probabilities can be associated with relatively high degradation levels while higher fault probabilities may occur at comparatively lower degradation levels. These findings demonstrate that fault risk and visual degradation provide complementary information and that their joint consideration can provide more comprehensive assessment of insulator condition for condition-based maintenance.

13:45
14:00

Paper ID: 106

Digital Learning Ecosystem in Ukrainian Economic and Management Education: Facts, Innovations, Prospects

Anastasiia Bilak; Tetyana Goran; Marta Dmytryshyn; Roman Rusyn; Antonina Tomashevska

Presentation: Online

Abstract: This paper investigates the functioning and integration of educational digital tools within the modern digital learning ecosystem, with a particular focus on Ukrainian economic and management education. To evaluate the main trends and user adoption patterns, an student poll was conducted, which made it possible to identify the core usage purposes such as presentations, data analysis, calculations, and language learning alongside interface preferences. Furthermore, the technical, pedagogical, and security characteristics of leading educational platforms were systematically compared. A SWOT analysis outlined strategic pathways, and a bimatrix game-theoretic model simulated interactions between platform developers and student users, revealing optimal equilibrium points. In addition, the application download trends on Google Play and the App Store were analyzed and their forecast was made by a sixth-order branched continued fraction model. The research findings highlight the prospects for intensifying the digital educational tools use in crisis conditions, emphasizing the integration of artificial intelligence, personalized learning, and digital literacy. Finally, the study provides an effective tips for educators, developers, and institutional decision-makers aiming to optimize resource management and sustain the quality of higher education digitally.

14:00
14:15

Paper ID: 91

Utility Is Not Compliance: Verifying Federated Unlearning for Autoencoder-Based Anomaly Detection

Oluwapelumi Alagbe

Presentation: Online

Abstract: Abstract. GDPR Article 17 requires that a participant’s influence be removable from a trained model. Federated unlearning (FU) is how that removal is attempted, but the available methods were built and tested on classifiers. We ask whether they survive transfer to a reconstructive model, using federated convolutional autoencoders for anomaly detec_xFFFE_tion on multi-site brain MRI (MICCAI FeTS 2022) across two forgotten_xFFFE_client regimes and three seeds. Our evaluation separates two things the literature usually reports as one: utility, meaning downstream detection performance, and fidelity, meaning distributional closeness to an oracle retrained without the forgotten client. Only the second speaks to com_xFFFE_pliance, and we find the two can be anti-correlated. Gradient ascent is the only method whose fidelity improves reliably. Knowledge distillation beats the oracle on utility in both regimes while its fidelity collapses as the forgotten client shrinks, and we trace this to its push-away objective, which diverges from the teacher even when the teacher already matches the oracle. Deleting that objective outright improves fidelity by 63% and 82%, several times a directly measured noise floor. FedEraser does not transfer at all. Reconstruction-loss membership inference turns out to be useless as a check: it already reports chance-level attack success on the pre-unlearning model, so it would certify a model that has forgotten nothing. Code, configurations and tests are released.

Keywords: Federated Unlearning · Trustworthy AI · Right to be Forgotten · Anomaly Detection · Autoencoders · Reproducibility.

14:15
14:30

Paper ID: 21

A Machine Learning based Multistage Stochastic Programming Framework for South African Retirement Portfolios

Charney Akala

Abstract: Black South African entrepreneurs experience retirement planning challenges stemming from their low financial literacy and volatile income streams excluding them from formal retirement systems. Existing South African retirement portfolio literature overlooks this demographic relying on static Mean Variance Optimization (MVO) frameworks and regression models that ignore uncertainty, sequencing risks, and rebalancing, leaving entrepreneurial portfolios mis-specified and inadequate for retirement adequacy. Aim: The paper aims to develop and validate a Machine Learning (ML) based Multistage Stochastic Programming (MSP) framework to construct retirement portfolios suited to the financial realities of Black South African entrepreneurs. Setting: The study is situated within the South African retirement landscape employing financial and macroeconomic data from January 2004 to January 2024 to construct retirement portfolios. Method: A simulation-based causal-comparative design relied on monthly index returns from eight asset classes which facilitated the calibration of the MSP framework, validating its performance relative to the MVO framework and static benchmarks using risk-adjusted and downside metrics.

Results The MSP framework consistently outperformed the MVO framework and static benchmarks across performance metrics displaying statistically significant excess returns and robustness thorough market regimes attributable to lower domestic risk-on factor exposure. Conclusion: The MSP framework provides a superior alternative to static optimisation frameworks for Black South African entrepreneurs addressing the specific retirement vulnerabilities of this population. Contribution: The paper contributes a fully integrated MSP framework which tailor made for Black entrepreneurs. This framework addresses demographic-specific and dynamic optimization gaps in South African retirement planning literature.

14:30
14:45

Paper ID: 11

Situational Awareness as a Legal Function in AI-Enabled Airport Security Governance

Kateryna Vodolaskova; Olena MAKEIEVA; Viktoriya CHEREVATIUK; Liliia NEVARA

Presentation: In Person

Abstract: The increasing integration of artificial intelligence (AI) into airport security governance is transforming the processes through which operational information is generated, interpreted and used for decision-making. Traditional aviation legal frameworks have largely been based on human-centred models of perception, judgement and responsibility. However, AI-enabled environments increasingly rely on interactions between human operators and technological systems, raising questions regarding the legal significance of situational awareness. This article examines whether situational awareness should be understood not only as a cognitive capability of individual actors but also as a legally relevant operational function within AI-enabled airport security governance. The study combines doctrinal legal analysis with an interdisciplinary conceptual approach based on situational awareness theory, distributed cognition and socio-technical systems research. The article argues that, as AI systems increasingly contribute to the collection, processing and prioritisation of operational information, situational awareness becomes a distributed function generated through interactions between human and technological actors. It further demonstrates that existing accountability models, traditionally centred on identifiable human decision-makers, require reconsideration in technologically mediated environments. The article proposes a distributed accountability approach that recognises the role of technological infrastructures in shaping the conditions of lawful and safe decision-making.

14:45
15:00

Paper ID: 22

Implications of Conducive Learning Environment as a form of Expanded Learning Opportunity in Outcomes-Based Teaching and Learning: An Assessment of the Learning Modality Preference of University Students in General Education Courses

Geoffrey Rhoel Cruz

Abstract: In response to the calls raised the Commission on Higher Education, the X University adopted the educational framework of Outcomes-Based Education since 2012 which provide for an expanded opportunity. Since then, the university has innovated its learning. Given the variety of learning modality implement by the university from 2021-2024, this study reassesses the effectiveness of a fully online mode of learning by assessing the feedback of students about their mode of learning preference. Using a quantitative analysis of data, findings showed that it is important for students to feel comfortable and confident with their learning environment for effective learning to take place. From the eight quarters surveyed spanning from three academic years, learners expressed their preference for the Full Online Mode of learning. Moreover, the findings affirmed that all identified course outcomes in Readings in Philippine History (GED103), a general education course, have been achieved using the Full Online Mode of learning. This study concludes that general education courses which are mostly lecture classes and does not require the use of any special facilities and laboratories should capitalize on the expanded learning opportunities that comes with the usage of a Full Online Mode of learning.

15:00
15:15

Paper ID: 137

Cooperative Guidance for Decentralized Multi-UAV 3D Mapping and Autonomous Navigation

Salahudden Salahudden; Arpita Nagpal; Parth Sailwal; Paanav Bansal; Anirban Roy; Rakesh Kumar

Abstract: Autonomous exploration and mapping of complex structures using multiple unmanned aerial vehicles (UAVs) requires effective coordination, safe motion planning, and efficient distribution of exploration tasks. This paper presents a decentralized three-UAV framework for cooperative structural exploration and three-dimensional(3D) mapping of a two-story building environment. The proposed framework combines cooperative exterior coverage, safety-aware motion planning, shared 3-D occupancy mapping, building-opening detection, and frontier-based interior exploration. A layered navigation strategy leverages an inflated 3-D costmap, geometric collision rechecking, and predictive inter-UAV collision avoidance, while dynamic work-stealing allows available UAVs to take over unfinished exploration tasks from other agents. A complementary sensor-fusion setting considers depth-lidar and noisy stereo measurements to explore mapping under sensing uncertainty. Experimental evaluation and controlled ablation studies show positive inter-UAV separation in the evaluated baseline trials and show measurable degradation when anticipatory costmap guidance is removed. The work-stealing study also demonstrates increased UAV engagement and spatial contribution with imbalanced workloads. The results indicate that a combination of decentralized task redistribution and layered safety-aware navigation offers a promising basis for cooperative structural exploration and mapping but also indicates the need for further validation on different environments and physical UAV platform.

15:15
15:30

Paper ID: 78

Model-Guided Residual Beamforming for Wideband Multi-RIS-Assisted THz MU-MISO Systems

Akil Hamied

Abstract: The Terahertz band of frequencies is considered as the most promising enabling technology for the 6G wireless networks due to its large spectrum resources and very-high data rate. Severe propagation loss, molecular absorption, and rapidly varying channels, however, make the problem of beamforming optimization in MISO systems with wideband multi-reconfigurable intelligent surface quite challenging, especially when accommodating multiple users. In this paper, we propose a Model-Guided Residual Beamforming framework based on the combination of analytical optimization and deep reinforcement learning for joint active and passive beamforming to cope with these challenges. A physically meaningful beamforming reference is set by the proposed framework using Regularized Zero-Forcing and Alternating Optimization, and bounded residual corrections for the digital Precoder and RIS phase configuration are learned by a Deterministic Deep Policy Gradient .The proposed formulation can reduce the effective optimization space, helps to stabilize the convergence of the policy learning process and maintains physically feasible beamforming solutions, by only allowing residual beamforming updates. The proposed scheme is tested on representative wideband THz channels and benchmarked with four baseline schemes. Numerical results show similar orders of spectral efficiency gain over the range of transmit power considered with significantly reduced online computational complexity compared to iterative optimization approaches. The observed performance gains are further validated by statistical analysis using paired hypothesis testing, bootstrapping and effect-size analysis. The proposed framework shows that it offers efficiency, robust, and scalable optimization solution for future wideband multi-RIS-assisted THz wireless networks.

15:30
15:45

Paper ID: 45

The Legibility Criterion for Recursive Self-Revision in Retinal Image Decision-Making

Rami Mohd Alroobi; Hong Cheng; Pooja Yogesh Athawale

Abstract: Recursive self-revision, in which a large language model (LLM) repeatedly evaluates and modifies decisions derived from a frozen model, can demand substantial implementation and evaluation effort. We propose a low-cost pre-flight diagnostic to estimate its utility, motivated by two negative results in retinal imaging. The diagnostic is based on a legibility criterion: self-revision is unlikely to help unless the failure signal is both legible, or interpretable at the individual-case level, and non-redundant, or not already exploited by a simple baseline. In Study A, an LLM revises a strategy memo for diabetic-retinopathy grading on IDRiD. The LLM is the weakest of four proposal mechanisms, and continued revision does not outperform a single revision. The optimization objective is effectively saturated, leaving no sufficiently legible signal for further revision. In Study B, which examines retinal referral on APTOS, the diagnostic returns NO-GO because a calibrated confidence threshold performs comparably to a learned rejector that also uses image-quality features. Here, the signal is legible but largely redundant. Tests on controlled data and real degraded images show graded sensitivity to planted non-redundant structure. The diagnostic returns NO-GO when a detectable confound adds no headroom over the baseline or when the available feature duplicates confidence, and it returns GO when real-image degradation introduces exploitable structure. The principal contribution is a screening procedure for assessing recursive self-revision before incurring its implementation cost. These null findings motivate and evaluate the procedure; they do not establish that recursive self-revision is ineffective in general.

Technical Session - 10 Online Time: 13:30 - 15:45. Location: Virtual Room - 2

Join Session

Session Chairs: Mr. MD KAVIUL HOSSAIN, Dr. Yüksel Çelik

13:30
13:45

Paper ID: 123

Strategy Map: Making Procedural Knowledge Explicit and Inspectable for Tool-Using Agents

Jisoo Jang; Wen-Syan Li

Presentation: In Person

Abstract: Tool-using language agents need procedural knowledge about required evidence, intermediate dependencies, and decision prerequisites that tool descriptions may omit. This paper specifies and evaluates Strategy Map, a milestone graph that makes these relationships explicit and inspectable. A reusable Published Map defines the procedure, while a session-local Working Map records values from execution traces. Additional context and model-mediated updates keep procedural requirements and accumulated evidence available to guide subsequent tool use. With this richer execution context, mean accuracy across three reasoning configurations on seven sales and operations planning workflows increases from 42.6% to 65.5% with Reasoning and Acting (ReAct) and from 28.3% to 66.9% with a retrieval-assisted executor. On FlowBench, mean completion improves in all four executor/scope comparisons, and scenario analysis identifies gains accompanied by recovery of task-relevant actions. A trace audit further shows how an answer can be produced before a prescribed decision tool is invoked. The study contributes an explicit separation of procedural requirements from execution-specific evidence and an empirical analysis of how procedural guidance relates to task completion and decision-step omission. Together, these contributions provide a basis for guiding tool-using agents and making their execution evidence accessible for human inspection.

13:45
14:00

Paper ID: 87

Data-Efficient Projection-Related Tuning of BioMedCLIP for Ultrasound Image–Text Retrieval Using US-365K

Tung Bui; Thanh Ma; Chau Tran

Presentation: Online

Abstract: Purpose: Ultrasound image–text retrieval requires alignment between sonographic appearance and textual descriptions under substantial visual variability and semantic redundancy. This study evaluates whether BioMedCLIP can be adapted to ultrasound using a constrained set of projection-related parameters rather than full-model finetuning, and whether improvements observed on a controlled test subset persist when retrieval is evaluated against the complete US-365K test gallery. Approach: We adapt BioMedCLIP using the public US-365K dataset. Most pretrained parameters are frozen; trainable parameters are selected by implementation-level name matching for proj or projection, together with the learnable logit_scale. Data-efficiency experiments use 1k, 5k, 10k, and 50k training subsets with fixed 5k validation and test subsets. The selected FT-50k checkpoint is additionally evaluated on the complete usable test split, leaving 71,918 valid image–text pairs after exclusion of one unreadable image. Results: On the controlled 5k test subset, FT-50k increases I2T Recall@10 from 3.88% to 7.60% and T2I Recall@10 from 4.46% to 7.90%. On the 71,918-pair full gallery, I2T Recall@10 increases from 0.3740% to 0.4436% (p = 0.0217), whereas T2I Recall@10 increases from 0.4602% to 0.6938% (p < 0.001). A complementary strict-greater rank statistic decreases from 5,806 to 2,536 positions for I2T and from 5,631 to 2,482 positions for T2I. Conclusions: Projection-related tuning of a largely frozen BioMedCLIP model produces measurable ultrasound-specific adaptation using 50,000 training pairs and preserves a ranking advantage when retrieval is expanded to the 71,918-item gallery. Clinical task performance was not evaluated and requires separate validation.

14:00
14:15

Paper ID: 71

System Reliability of Airport Security Screening Under Varying Operational Conditions

Pavel Juhasz; Zbyšek Korecki; Tomáš Hoika

Presentation: In Person

Abstract: Airport security screening represents a critical safety function in civil aviation, where system reliability depends on both technological capability and human performance. This study evaluates the reliability of passenger and baggage security by quantifying task-level human error and assessing its impact on overall system performance. Security screening was analysed as a multi-layer system composed of sequential human-operated detection tasks. The Human Error Assessment and Reduction Technique was applied to quantify Human Error Probability (HEP) for seven key screening tasks identified through operational observation at an international airport. Each task was classified according to Generic Task Type (GTT), and relevant Error Producing Con-ditions (EPCs) were identified to calculate task-specific HEP values under peak operational load. System-level reliability was subsequently modelled by combining individual task error probabilities and evaluating multiple operational scenarios with varying staffing levels. The analysis revealed substantial variability in human error probability across individual screening tasks. The highest HEP was identified in X-ray image analysis (22.68%), while routine procedural tasks such as handheld metal detector screening exhibited considerably lower error probabilities (0.45%). Despite these task-level dif-ferences, the multi-layer architecture of the screening process resulted in extremely high system reliability under standard operational conditions (HEP ≈ 9.9 × 10⁻⁷). Sce-nario analysis further demonstrated that reduced staffing significantly increases system failure probability, reaching 1.01% when all screening tasks are performed by a single operator.

14:15
14:30

Paper ID: 104

On the Usage of Computer Vision Algorithms, Deep Learning, and Large Language Models for Feature Extraction from House-Tree-Person (HTP) Sketches

Abdul Samad --; Dila Gurer; Elena Battini Sonmez; Sibel Halfon

Abstract: TheHouse–Tree–Person (HTP) technique is a projective drawing based assessment in which individuals are asked to draw a house, a tree, and a person. Interpretation typically considers both the content of the HTP sketches and specific drawing features, including size, placement, proportions, line quality, level of detail, omissions, and the spatial rela tionships among depicted elements. The presence/absence of particular traits allows to infer the level of anxiety, depression, and aggressive be haviors in the person taking the test. The aim of this research is to auto mate the analysis of HTP drawings, starting with algorithms that extract relevant features from HTP sketches. Feature extraction has been done using Computer vision (CV), Deep learning (DL), and Large Language Model (LLMs) techniques, and their performance has been compared. Deep Learning (DL) and Large Language Models have achieved the best results, correctly classifying most HTP features. Future work will focus on expanding the clinical dataset and using more advanced deep learning models for improving classification performance.

14:30
14:45

Paper ID: 119

Physics-Informed Neural Networks for Modeling Athero-sclerotic Inflammation: A Quantitative Study of a One-Dimensional Reaction-Diffusion System

ahlam Elkhettab; Imane Elberrai

Abstract: Atherosclerosis is a chronic inflammatory disease in which lipid accumulation and immune cell recruitment contribute to arterial plaque formation and progression. This study evaluates a Physics-Informed Neural Network (PINN) formulation of an established one-dimensional reaction-diffusion model describing the early interaction between recruited immune cells and proinflammatory cytokines. The objective is computational rather than clinical: to determine whether a standard PINN can accurately reconstruct the spatiotemporal solution while satisfying the governing partial differential equations, initial conditions, and no-flux boundary conditions. This study uses no patient or experimental dataset. An independent finite-difference method (FDM) solution provides a synthetic numerical benchmark, while disjoint validation and test points are reserved for model selection and final assessment. On 4,000 held-out test points, the PINN achieved relative L2 errors of 8.60×10⁻4 for the immune-cell field M and 3.17×10⁻3 for the cytokine field A, with R2 scores of 0.999995 and 0.999949, respectively. Line-profile comparisons, absolute-error maps, and convergence curves confirm close agreement with the FDM reference. The present work is restricted to forward simulation; inverse parameter identification and patient-specific inference are identified as future extensions rather than demonstrated outcomes.

14:45
15:00

Paper ID: 138

Selection of Psychological Features for the House-Tree-Person (HTP) Test

Abdul Samad; Dila Gurer; Elena Battini Sonmez; Sibel Halfon

Abstract: TheHouse-Tree-Person (HTP) test is a projective test widely used in Psychology to assess personality-related issues. Feature selection is an important research area in machine learning that aims to reduce data dimensionality by discarding irrelevant and/or redundant features. This paper combines these two research fields by applying feature selection algorithms to all psychological features extracted from the sketches drawn during the HTP test. The final aim is to answer an open question in psychology, i.e., to identify the best subset of features that most strongly correlate with emotional difficulties. Several feature extraction algorithms have been run, and their subsets have been compared.The best performance of 81.32% was achieved by the Extra Trees classifier using the wrapper-based backward elimination technique with only seven features. The algorithm and the best subset of features have been identified. Future work will focus on the application of more advanced feature selection techniques to further improve model performance and identify the most relevant traits.

15:00
15:15

Paper ID: 172

Signature-Grounded Validation of Explanations for Smart Contract Vulnerability Detection

Muhammad Sajjad; Hammad Ali; Aamir Raza; Muhammad Atif Saeed; Azka Atiq; Akhtar Jamil; Alaa Ali Hameed

Abstract: Explanations of learned smart contract vulnerability detec- tors are usually judged by a deletion test, which accepts them if re- moving the highlighted instructions changes the prediction. We show that this test is not enough. On the VASCOT Transformer, attention- based explanations remove every verdict they explain on its verified- contract dataset. Yet on its deduplicated four-class dataset they place SELFDESTRUCT among the top 5 % of instructions in 81–100 % of con- tracts, whichever class is being explained, and they remove hardly more verdicts than a simple rarity ranking. To expose such failures we propose signature-grounded validation, which adds to deletion and sufficiency tests a class-specificity test and a counterfactual, both derived from the MAIAN definitions behind the labels. Signed attributions pass this vali- dation. With X-VASCOT, our class-contrastive adaptation of Integrated Gradients for windowed bytecode models, SELFDESTRUCT appears in the top 5 % of every suicidal explanation and of no safe or greedy one. An audit of the VASCOT datasets puts this in context: 88 % of test con- tracts in the two larger datasets have an exact copy in training, and on the verified-contract dataset accuracy on unseen contract families drops below the majority rate.

15:15
15:30

Paper ID: 25

Artificial Intelligence in Criminal Justice: a Legal Analysis of Veripol and RisCanvi

Alberto Jose de Nova Labián

Abstract: The increasing use of artificial intelligence within the criminal justice system has become a defining feature of contemporary European legal developments. This paper examines the implications of this transformation through two case studies from the Spanish context: the Veripol system, designed to detect false criminal complaints, and the RisCanvi programme, used to assess the risk of reoffending in the penitentiary sphere. By analysing their application at different stages of the criminal process, the study highlights the growing role of algorithmic decision-making in areas traditionally governed by human judgment. While these tools offer clear benefits in terms of efficiency and systematisation, their use raises important legal concerns. In particular, issues relating to reliability, transparency and the protection of fundamental rights emerge as central challenges. The analysis also identifies the risk of automation bias, whereby decision-makers attribute excessive weight to algorithmic outputs, potentially affecting the fairness of judicial and administrative decisions.The paper argues that the legitimacy of such systems depends on their integration within a rights-based framework that ensures meaningful human oversight, procedural safeguards and accountability. Ultimately, the findings suggest that technological innovation must remain compatible with the core principles of criminal law and the rule of law.

15:30
15:45

Paper ID: 140

Deep Learning-Based Adult Content Classification: A Comprehensive Performance Analysis of EfficientNetV2, ResNet50, and VGG19 on a Novel Multiethnic Adult Content Dataset

MD KAVIUL HOSSAIN; Md Foysal Ahmed; Md Ali Rishan; Rafsun Hossnain

Abstract: Adult content detection is a crucial task when it comes to the differentiation of nudity and genitalia exposure. Detecting explicit adult content solely based on skin exposure percentage is not efficient and effective since the illustration of genitalia is not considered in such calculations. And when it comes to capturing the hidden body parts of humans, machine and deep learning models struggle greatly. Keeping this in mind and combining the salient factors and parameters, we have developed a completely new and novel adult content dataset on our own which contains three distinct classes – Safe For Work (SFW), Not Safe For Work 1 {NSFW1) & Not Safe For Work 2 (NSFW2). These classes contain images of fully clothed people, partial nudity and genitalia exposure & complete nudity with full genitalia exposure, respectively. To conduct this research, we have taken an approach where we have iterated three separate models- EfficientNetV2, ResNet50 & VGG19 with their basic framework and later with advanced fine-tuning techniques such as data augmentation with random flips and rotations, Gaussian noise and blur augmentation, early stopping strategies, dynamic class weights and categorical focal loss measurements. The sole purpose behind such an approach is to analyze the relative performances of the models with and without framework consolidation. The obtained testing accuracies of conventional model frameworks are 98% in VGG19, 40.58% in EfficientNetV2 & 93.61% in ResNet50 while consolidated models produced testing accuracies of 92%, 15% & 76.07% in VGG19, EfficientNetV2 and ResNet50 respectively suggesting that VGG19 outperformed the other models in all scenarios

2nd International Conference on Applied Computational Intelligence and Systems Modeling