Prof. Alessandro Bogliolo, DiSPeA, University of Urbino, Italy
Alessandro Bogliolo is full Professor of Computer Systems at the University of Urbino, Italy. He received the Laurea degree in Electrical Engineering and the Ph.D. degree in Electrical Engineering and Computer Science from the University of Bologna, Italy. From 1992 to 1999 he was with the Department of Electronics and Computer Science (DEIS), University of Bologna. In 1995 and 1996 he was with the Computer Systems Laboratory (CSL), Stanford University, CA. From 1999 to 2002 he was with the Department of Engineering (DI), University of Ferrara, Italy. His research interests include mobile crowdsensing, machine learning, and digital platforms for participatory social innovation. In 2019 he co-founded Digit srl, benefit corporation for digital social innovation.
Speech Title: Beyond Benchmarks: A Participatory Framework for Real-World ML Dataset Creation
Abstract: The availability of representative datasets is increasingly shaping ML research and applications, because of the crucial role they play in model training and validation. Domain-specific annotated open datasets, released by research groups on the basis of their well documented and controlled experiments, have provided an enormous contribution to the advancement of ML in their respective application areas, acting as key enablers and attracting research efforts worldwide. Moreover, having many independent teams working on the same data makes it possible to conduct comparative assessment, which is a fundamental task in experimental studies. However, relying on a small number of datasets recognized as benchmarks within a community, risks to polarize research efforts, limit their scope, and make it difficult to transfer results to real-world applications operating under different conditions. On the other hand, building ad hoc in-house datasets tailored to the unique needs of a given application, would significantly reduce productivity and distract researchers from their primary goal, adding a cumbersome milestone along their research roadmap. Whenever a ML model is conceived to run on mobile/wearable devices equipped with sensors capable of sampling and feeding input features into the model, the same target devices and sensors could in principle be used to conduct preliminary experiments to build training/validation datasets. In this scenario, real-world crowdsensing could provide a viable complement to controlled lab experiments, trading off data quality for size, diversity, and transferability.We present an online GDPR-compliant framework that allows researchers to design their protocols and to coherently apply them to gather annotated data coming from lab experiments, synchronous mass acquisition sessions, and asynchronous crowdsensing campaigns. Data originated from all these sources coexist in the framework despite their different quality, but they can be filtered according to the needs of each project, based both on quality metrics and on metadata representing the experimental conditions and the reliability of the source. A real-world use case focused on human activity recognition is presented and discussed.
Prof. Hua-Liang Wei, University of Sheffield, UK
Dr Hua-Liang Wei received the Ph.D. degree in the Department of Automatic Control and Systems Engineering, the University of Sheffield, UK, in 2004. He currently works with the School of Electrical and Electronic Engineering, the University of Sheffield. Dr Wei is head of the laboratory of Dynamic Modelling, Data Mining and Decision Making (3DM), head of the laboratory of Digital Medicine & Computational Neuroscience (DMCN), and deputy head of the Solar Physics and Space Plasma Research Centre (RP2RC). He has been awarded grants (as PI and Co-I) from a variety of funding bodies including EPSRC, NERC, STFC, EU H2020, the Royal Society and so on. His main research interests are in signal processing, system identification and data-driven modelling of nonlinear complex systems, explainable artificial intelligence, interpretable machine learning, deep learning, and intelligent diagnosis, with applications in many interdisciplinary fields. He has been devoting to these areas for over 20 years, tackling challenging issues at the interface of Data Science, Computer Science and Control & Systems Engineering, to develop new and fundamental methods that can be adapted to solve challenging, complex systems modelling problems in multi- and inter-disciplinary areas including engineering and industry, seasonal weather and climate, space weather and space environment, biomedical and neuroscience. He has a strong interest in initialising and developing new research directions that can help solve emerging challenging problems in multidisciplinary domains. He has published over 170 peer-reviewed papers.
Assoc. Prof. Tuna Çakar, MEF University, Turkey
Dr. Tuna Çakar is an Associate Professor in the Department of Computer Engineering at MEF University’s Faculty of Engineering in Istanbul, where he teaches and supervises applied data science and machine/deep learning projects that connect computational methods with human behavior. His academic path is interdisciplinary: he completed his B.Sc. in Biological Sciences and Bioengineering at Sabancı University, earned an M.Sc. (thesis) in Cognitive Science at Boğaziçi University, and received his Ph.D. in Cognitive Science from Middle East Technical University (METU). At MEF University, Dr. Çakar’s research spans data analysis, machine learning, deep learning, algorithm development, and decision-making, with a particular emphasis on applied neuroscience and neuro-cognitive measurement. Across publications and collaborative projects, he investigates how behavioral, psychological, and neurophysiological signals can be transformed into actionable models—ranging from user- and consumer-level prediction tasks to decision support and risk analytics. He has contributed to studies that pair modern learning methods with neuroimaging and electrophysiology (e.g., fNIRS and EEG paradigms) to probe the neural correlates of judgments and choices, and he also works on real-world ML applications such as fraud detection, quality inspection, and customer analytics. He is passionate about translating academic research into practical tools and frequently collaborates with industry partners on projects involving customer analytics, fintech decision support, and human-centered AI. Dr. Çakar mentors undergraduate and graduate researchers and contributes to interdisciplinary initiatives spanning engineering, psychology, and neuroscience. In his courses, he emphasizes reproducible experimentation, responsible data use, and clear communication of model assumptions and limitations to stakeholders. (ORCID: 0000-0001-8594-7399.
Prof. Meng Wang, National Science Library, Chinese Academy of Sciences, China
Meng Wang is an outstanding postdoctoral researcher at the National Science Library, Chinese Academy of Sciences. His research focuses on data mining, knowledge graphs, and large language model reasoning. He has published over 10 papers as first author in CCF, JCDL, SCI, EI-indexed journals, and core journals. He has been awarded the "Top Academic Paper Award of China's Premium Science and Technology Journals." He has participated in and chaired sessions at international conferences on 5 occasions. He holds over 20 authorized invention patents and software copyrights as first author. He currently serves as a reviewer for *Data Analysis and Knowledge Discovery* and *Data Intelligence*.
Speech Title: Semantically-Guided Two-Stage Classification: Integrating Structural Awareness and Expert Routing
Abstract: As the pace of scientific knowledge production accelerates and the volume of domain-specific literature continues to grow, the development of classification models capable of capturing semantic subtlety and adapting to imbalanced label distributions has become increasingly central to intelligent text understanding. However, existing multi-label classification approaches for scientific texts often fail to adequately leverage semantic features and struggle with accurately identifying low-resource or semantically overlapping categories. To address these challenges, this study proposes a two-stage classification framework that integrates domain-aware semantic feature encoding into large language models (LLMs). In the first stage, we construct a dynamic-window-based topic semantic extractor and a hierarchical structure-aware encoder, embedding these features into the [CLS] and [SEP] positions of the LLM to enhance semantic representation. In the second stage, a fine-grained semantic routing strategy is introduced within a mixture-of-experts (MoE) architecture, enabling adaptive expert allocation informed by semantic cues. Experimental results on both the DBpedia benchmark and a domain-specific dataset of scientific value sentences demonstrate consistent performance improvements: our approach achieves a 4.37%–5.82% gain in F1 score over existing semantic encoding methods, reaches a peak F1 of 94.19% on value sentence recognition, improves macro-average F1 to 93.35% on balanced data (a 9.04% increase), and boosts F1 for minority classes by over 25% on imbalanced datasets. The results suggest that semantically guided representation enhancement may offer a promising pathway toward more accurate and resilient classification in scientific text mining.
Assoc. Prof. Dr. Tuğçe Çelik, OSTIM Technical University, Turkey
Assoc. Prof. Dr. Tuğçe Çelik is an architect and Associate Professor at the Faculty of Architecture and Design, OSTİM Technical University, Ankara, Türkiye. She conducts her research at the intersection of architecture and generative artificial intelligence. Her work critically examines how generative AI systems (text-to-image and multimodal models) function as design tools that produce, transform, and reinterpret architectural knowledge, spatial typologies, and cultural patterns. Grounded in an architectural design–oriented academic background, Dr. Çelik develops interdisciplinary evaluation frameworks that integrate architectural reasoning with computational and perceptual methods, including performance-based simulations (e.g., daylight analysis). Through these frameworks, her research addresses contemporary challenges emerging at the intersection of architecture, design cognition, and human–AI co-creation, with a particular focus on concepts such as spatial logic, coherence, and environmental performance in generative design outputs. By positioning generative artificial intelligence not as a purely technical system but as a mediating design instrument, her work contributes to interdisciplinary debates on AI in design, creative authorship, and the epistemology of AI-assisted architectural production. Dr. Çelik’s research has been published in internationally indexed journals (Web of Science and Scopus) and presented at interdisciplinary international conferences. Within the scope of international academic collaboration and mobility, she has taught at VIKO – Vilniaus Kolegija under the Erasmus+ Staff Mobility (Teaching) program and participated in Erasmus+ Staff Mobility (Training) programs at institutions such as Universidad de Sevilla and Univerzitet u Beogradu. She currently serves as a coordinator of Erasmus+ projects focusing on artificial intelligence, architecture, and sustainable design.
Assoc. Prof. Marko Đurasević, University of Zagreb Faculty of Electrical Engineering and Computing, Croatia
Marko Đurasević is an Associate Professor at the Faculty of Electrical Engineering and Computing (FER), University of Zagreb. His research is centered on evolutionary computation, particularly genetic programming and hyper-heuristics for solving complex scheduling and optimization problems. He earned his Ph.D. in Computer Science from FER in 2018, with a dissertation focused on the automated design of dispatching rules in unrelated machines environments. Dr. Đurasević has published over 100 scientific papers in international journals and conferences, contributing extensively to the fields of combinatorial optimization, machine learning, and soft computing. He is the principal investigator of two nationally funded projects dealing with optimization of containers in ports and routing of electric vehicles. Furthermore, he also leads a project in collaboration with the company AVL-AST. His scientific excellence has been recognized with the Annual Award for Young Researchers by the Croatian Parliament in 2023 and several other national institutions. Dr. Đurasević is an active member of IEEE, IEEE CIS, ACM, and ACM SIGEVO, and regularly serves as a reviewer for leading journals in artificial intelligence and operations research.
Speech Title: Using Genetic Programming to Automatically Design Heuristics for the Electric Vehicle Routing Problem
Abstract: The vehicle routing problem (VRP) is a famous combinatorial optimization problem in which the goal is to determine the best routes for a given number of vehicles, such that all the customers are served, and that one or more user defined criteria are optimised. Due to the growing environmental concerns that arise due to the negative influence of people on the environment, the electric VRP (EVRP) variant became increasingly popular and extensively researched in the literature. However, obtaining solutions for large or dynamic problems is difficult, as not all information about the problem is known beforehand. Therefore, it is required to use simple constructive heuristics methods, which incrementally construct the solution to such problems. These heuristics can quickly react to the changing conditions of the problem by taking into account the latest available information. Since there are numerous problem variants and criteria that need to be considered, it is difficult to design heuristics manually. Therefore, genetic programming (GP) has been used to automatically design heuristics for a wide range of problems. In the context of designing heuristics, GP can be used to create an arbitrary expression used rank all the decisions in the system. Based on these rankings the heuristics selects the appropriate decision and executes it. This possibility of designing an arbitrary expression to rank the decisions that need to be performed allows GP to design heuristics for almost any problem variant or optimization criterion that is considered, thus relieving human experts of the tedious task of manually designing heuristics. The initial results demonstrate that by using such an approach it is possible to achieve improved performance in comparison to manually designed rules for EVRP.
Dr Hanyu Li, University of Chinese Academy of Science；National Science Library, Chinese Academy of Sciences, China
Hanyu Li is Senior Engineer, Specially Appointed Backbone Researcher (CAS), and Associate Supervisor of Doctoral Students. Currently serving as Deputy Director of the Knowledge Systems Department at the National Science Library, CAS, Hanyu Li specializes in intelligent knowledge services, informatics, and trustworthy digital scholarship. She has spearheaded more than 20 major research and infrastructure projects, notably overseeing sub-tasks of the National Key R&D Program and CAS “14th Five-Year Plan” strategic initiatives. Her portfolio also encompasses numerous youth talent development and institutional construction programs.
Assoc. Prof. Goran Oreški, University of Pula, Croatia
Assoc. Prof. Dr. Goran Oreški is an Associate Professor at the University of Pula, Croatia, and Head of FIPUlab. He received his PhD degree from the University of Zagreb, where his research focused on data analysis and machine learning. His academic work focuses on machine learning and artificial intelligence, with a primary emphasis on learning from structured and tabular data. His research addresses methodological aspects of data-driven modeling as well as its application in complex and data-intensive environments. Dr. Oreški leads a research team within a laboratory focused on computer vision, where his work includes the development and application of visual data analysis methods for applied research problems. Alongside this, he has extensive experience in the design and implementation of data processing pipelines and intelligent systems, and has led and participated in multiple scientific and industry-oriented projects. His research has been published in internationally indexed scientific journals and presented at international conferences. Through his academic and project activities, he contributes to the advancement of applied machine learning and artificial intelligence.
Assoc. Prof. Kazuya Ueki, Meisei University, Japan
He received his B.S. degree in Information Engineering in 1997 and his M.S. degree in Computer and Mathematical Sciences in 1999, both from Tohoku University, Sendai, Japan. In 1999, he joined NEC Soft, Ltd., Tokyo, Japan, where he was mainly engaged in research on face recognition. He received his Ph.D. degree from the Graduate School of Science and Engineering, Waseda University, Tokyo, Japan, in 2007. From 2013 to 2017, he served as an Assistant Professor at Waseda University. He is currently an Associate Professor in the School of Information Science, Meisei University. His research interests include information retrieval, video anomaly detection, pattern recognition, and machine learning. He is involved in the video retrieval evaluation benchmark (TRECVID) sponsored by the National Institute of Standards and Technology (NIST), contributing to the development of video retrieval technology. His submitted systems achieved the highest performance in the TRECVID AVS task in 2016, 2017, 2022, and 2025.
Speech Title: From Matching to Reasoning: Vision-Language Models for Adaptive Video Retrieval
Abstract: The rapid evolution of vision-language models (VLMs) is reshaping how video retrieval systems interpret and respond to user queries. Rather than relying on direct keyword matching between text and visual content, recent approaches exploit multimodal reasoning and generation to enable more flexible and adaptive search behavior. This talk introduces four complementary strategies that illustrate this paradigm shift. First, query expansion at inference time uses large language models and image generation to create semantically diverse query variants, improving coverage of complex visual concepts. Second, retrieval models are adapted through caption-driven fine-tuning based on automatically generated descriptions, enabling effective domain customization without costly manual labeling. Third, a semantic verification stage based on visual question answering is applied to re-rank retrieved results, ensuring consistency between the query intent and visual evidence. Finally, an interactive retrieval framework integrates real-time vision-language reasoning, allowing users to iteratively refine results through follow-up questions or image-based feedback. Collectively, these techniques demonstrate how multimodal intelligence can transform video retrieval into a context-aware, user-adaptive process. Experimental observations, practical design considerations, and future research directions toward interactive video understanding will also be discussed.
