Prof. Anne Schwerk
IU International University of Applied Sciences
Anne Schwerk is Professor of
Artificial Intelligence at IU International University of Applied
Sciences and Deputy Department Head for IT & Engineering. In
parallel, she is a Senior Manager for Trustworthy AI at Accenture,
advising on AI governance, risk, and compliance (e.g., EU AI Act)
and leading cross-functional initiatives on responsible AI. Her
work centers on explainable and trustworthy AI, human-centered
evaluation, LLMs and NLP, and data quality for high-stakes
applications. She serves on the board of the German Data Science
Society (GDS) and speaks regularly at international forums on
explainable AI and AI in healthcare. Previously, Anne led the
TBase rollout at Charité/BIH, headed scientific management at the
Berlin Institute of Health Center for Regenerative Therapies
(BCRT), and held AI leadership roles at CENTOGENE and the German
Research Center for AI.
Speech title: "SHAP-Guided
Feature Reduction to Improve Cross-Subject Generalization in
Multimodal Physiological Emotion Recognition
Abstract: Multimodal emotion recognition from physiological signals has reported high classification performance in controlled evaluation settings, yet such results often rely on subject-dependent validation protocols that substantially overestimate real-world generalization. In this work, we systematically investigate whether explainability-driven feature reduction can improve subject-independent performance in multimodal physiological emotion recognition under strict leave-one-subject-out (LOSO) validation. Using the WESAD dataset, we study a three-class classification task (baseline, stress, amusement) based on chest-worn physiological signals and evaluate three representative model architectures: Random Forest, XGBoost, and a one-dimensional convolutional neural network. Feature relevance is quantified using SHAP values computed exclusively on training data within each LOSO fold. The resulting SHAP-based rankings are used to derive compact feature or channel subsets, followed by full model retraining under identical conditions. Across all evaluated architectures, SHAP-guided feature reduction consistently improves cross-subject generalization, supported by stable feature relevance across subjects. In particular, XGBoost trained on a reduced SHAP-selected feature set achieves statistically significant gains in both accuracy and macro F1-score compared to full-feature baselines. Class-wise analysis further reveals pronounced improvements for the amusement class, which is typically challenging to distinguish in physiological emotion recognition. These findings demonstrate that explainability can serve not only as a post-hoc analysis tool, but also as a principled mechanism for model optimization under realistic evaluation constraints. More broadly, the results highlight the importance of combining subject-independent validation with explainability-driven feature selection to obtain reliable and deployable physiological emotion recognition systems.
Prof. Silke Vaas
IU International University of Applied Sciences
Silke Vaas is a Professor of
Business Intelligence and Head of the Master’s degree programs in
Business Intelligence at IU International University of Applied
Sciences. Her work addresses research and practice in Business
Intelligence architectures and methods, Business Analytics,
Machine Learning, Intelligent and Multi-Agent Systems, affective
computing and large language models (LLMs). She brings over 15
years of industry experience as an IT project manager, where she
led numerous IT initiatives and was responsible for the
development, operation, and continuous advancement of more than a
dozen mission-critical operational, planning and Business
intelligence systems across the logistics, retail, finance, and
tax sectors. She earned her PhD at CIT in Ireland, focusing on
intelligent, learning user interface agents in affective
computing. Alongside her doctoral research, she led a research
project at the Institute for New Media in Frankfurt, strengthening
her expertise in applied AI and interactive systems.
Speech title:
"Aligning Code LLMs for Automated Unit Test Generation: A
Comparative Study of PPO and DPO
Abstract: Large language models (LLMs) can generate unit tests, but the resulting test suites often fail to execute, omit meaningful assertions, or achieve only superficial coverage without strong fault detection. We study alignment methods for unit test generation using a deterministic, execution-based oracle implemented as a robust reward engine. The oracle evaluates (i) compilability, (ii) executability, (iii) functional correctness against a canonical solution, and (iv) test quality via coverage and mutation testing. Starting from a supervised fine-tuned (SFT) code model as the baseline, we compare online reinforcement learning with Proximal Policy Optimization (PPO) to offline preference optimization with Direct Preference Optimization (DPO). We introduce a gated Staircase Reward that reduces reward sparsity by providing hierarchically structured and progressively refined feedback, ranging from syntactic validity to robust fault detection as measured by mutation testing. For DPO, we construct an automatic preference dataset via Synthetic Elitism by pairing elite candidates with weak or failing candidates under a reward margin. On TestEval, PPO and DPO improve execution correctness and coverage compared to SFT, with DPO achieving the strongest overall unit-testing performance. We quantify the alignment-induced performance trade-off on HumanEval+ and evaluate generalization on the contamination- free UnLeakedTestBench (ULT). Finally, we report computational costs to highlight DPO’s efficiency under the same oracle.
Asst. Prof. Peng Liu
Singapore Management University
Peng Liu has a Ph.D. in Statistics and Data Science and a M.S. in Business Analytics from the National University of Singapore. Currently he is an Assistant Professor of Quantitative Finance (Practice) at the Lee Kong Chian School of Business, Singapore Management University. He is also an adjunct research fellow at the Institute of Operations Research and Analytics at National University of Singapore. He has over ten years of industry experience across multiple industries. His research highlights expertise in areas such as deep learning, sparse estimation, and Bayesian optimization with applications in Finance.
Dr. Nuri Park
Hanyang University
Nuri Park is currently a
Postdoctoral Researcher at Hanyang University. She received a
Master's degree and a Ph.D. in Smart City Engineering from the
Hanyang University in Korea. She is interested in Traffic Safety,
Transportation Engineering, Data Mining Applications / Statistical
Analysis in Transportation, and Big Data Analysis in
Transportation & Logistics systems.
Speech title:
"AI-Driven Real-Time Traffic Crash Risk Prediction Using
Transportation Big Data
Abstract: To overcome the inherent limitations of roadway crash data, specifically its rarity, randomness, and heterogeneity, previous studies propose a proactive real-time crash risk prediction framework leveraging Big Data and advanced Artificial Intelligence (AI). This study introduces research trends in AI-driven real-time traffic crash risk prediction using transportation Big Data and develops a crash risk prediction model for Korean expressways. The proposed model is based on a Conditional Generative Adversarial Network (CGAN) and various machine learning classification models to enhance predictive accuracy and address data imbalances. The methodology is structured in three integrated stages: first, K-means Clustering (KC) is employed to define heterogeneous crash risk situations, followed by the identification of key precursors from traffic data using the Boruta-SHAP. Second, to address the critical issue of imbalanced datasets, a Conditional Generative Adversarial Network (CGAN) is utilized to augment crash samples while preserving the unique characteristics of each risk cluster. Finally, a crash risk classification model is developed.
Asst. Prof. Ayaz H. Khan
King Fahd University of Petroleum and Minerals (KFUPM)
Dr. Ayaz H. Khan is an Assistant Professor in the Computer Engineering Department at King Fahd University of Petroleum and Minerals (KFUPM), Saudi Arabia, and a Research Fellow at the SDAIA–KFUPM Joint Research Center for Artificial Intelligence. His research focuses on GPU-accelerated computing, distributed systems, high-performance computing, and AI-driven data analytics. He works on scalable AI infrastructure, multi-GPU optimization, distributed communication middleware, and energy-efficient computing systems. Dr. Khan is a certified NVIDIA Deep Learning Institute (DLI) Instructor and actively collaborates with industry and research partners on AI and high-performance computing solutions. His work contributes to building scalable and sustainable intelligent systems for next-generation computing applications. Beyond his academic and professional pursuits, he is an enthusiastic traveler, having visited more than 176 cities across 11 countries.
Speech title: "
Abstract: After natural disasters such as earthquakes, floods, or wars occur, cellular communication networks often sustain significant damage or become impaired. In these critical situations, first responders must coordinate with other rescue teams to communicate essential information to central command and survivors. To address this challenge, we have developed a reliable and rapidly deployable wireless ad hoc system for post-disaster management using Data Distribution Service (DDS) middleware, specifically RTI-DDS, named R-RDSP. The R-RDSP further enhances these metrics, achieving a 14.5% improvement in end-to-end delay and a 20.24% improvement in round-trip delay over the RDSP scheme. The R-RDSP system consists of three main modules: client, relay, and server. Each module connects to others via an ad hoc network, ensuring direct device-to-device communication without relying on existing infrastructure. The client module collects and sends the victim’s location and emergency messages. The relay modules forward these messages across the ad hoc networks, ensuring minimal delay and high reliability. Finally, the server module receives the messages, processes them, and coordinates the response. Leveraging RTI-DDS for reliable message distribution, the system demonstrates robust performance even under challenging network conditions.