Abstract. LLM agents are increasingly used to study opinion formation, information diffusion, and collective behavior. Yet a consequential layer of these simulations often remains implicit: agents do not respond to an interaction network itself, but to the information made visible in their contexts. The same agents and network may therefore produce different outcomes under different exposure mechanisms. This tutorial presents an exposure-centred view of LLM social simulation, tracing the path from potential interactions to visible evidence, agent updates, and collective outcomes. Communication networks and memory determine what information reaches each agent, while roles, influence, and other structural cues shape how that information is interpreted. Interaction traces and emergent network structures, in turn, reveal how local updates accumulate into collective dynamics. By treating these components as a continuous process rather than isolated design choices, the tutorial highlights structure as part of the mechanism that produces social behavior. This perspective supports more deliberate simulator design, clearer comparisons across systems, and more reliable conclusions about LLM-based social dynamics.

Bio. Dr. Chenhao Ma is an Assistant Professor at the Chinese University of Hong Kong, Shenzhen. Prior to that, he was a Postdoctoral Fellow at the University of Hong Kong. He received his PhD degree from Department of Computer Science in the University of Hong Kong (HKU) in 2021, advised by Prof. Reynold Cheng. He once was a visiting student in the University of New South Wales (UNSW) in 2019, working with Prof. Xuemin Lin. Till now, he has published more than 40 papers in the areas of database and data mining, including one of four Best of SIGMOD2020 (a world flagship conference in database areas, 4/458), and most of them were published in top-tier conferences and journals (e.g., SIGMOD, PVLDB, KDD, NeurIPS, ICML, ICLR, VLDBJ, TKDE, and TODS). He was awarded the ACM SIGMOD Research Highlight Award 2021. He has served as area chairs / PC members and reviewers for several top conferences and journals (e.g., VLDB, ICDE, KDD, WWW, CIKM, TKDE, and VLDBJ).

Abstract. AI systems used in behavioural and social computing expose attack surfaces that go far beyond small pixel perturbations. Adversaries may jointly manipulate social-graph links and user features, rewrite prompts, coordinate text-image attacks, poison retrieved documents, or influence the tool returns consumed by an AI agent. This tutorial develops a unified view of adversarial and certified robustness for these settings. Starting from classical adversarial examples, Lipschitz methods, and randomized smoothing, it moves to social graphs, language and multimodal models, and tool-using agents. A central theme is the separability barrier: guarantees for individual channels do not necessarily compose when those channels interact. The tutorial closes with practical limitations and open research problems for trustworthy behavioural and social AI.

Bio. Blaise Delattre is a JSPS Postdoctoral Fellow at the Institute of Science Tokyo, where he works with Prof. Yang Cao. His research focuses on certified robustness, with contributions to Lipschitz-constrained neural networks, randomized smoothing, and robustness guarantees for multimodal and foundation models. He received his PhD in Computer Science from Université Paris-Dauphine PSL in 2025. His work has appeared at ICML, ICLR, AISTATS, ACL, and ECAI, including oral and spotlight papers.

Abstract. Community search is a key graph query task that finds cohesive communities based on queries, with uses in social network analysis, team building, recommendations, bioinformatics, and fraud detection. Early work centered on dense subgraph models, efficient algorithms, and offline indexing. Recently, research has expanded to include learning-based methods, complex networks, and interactive querying. This tutorial offers a systematic overview of community search in graphs, summarizing foundations, current advances, and future directions. We first present common community models and main algorithms, such as exact and heuristic solutions, pruning, and indexing. Next, we discuss learning-based approaches, including graph neural networks and transformers, and compare them to traditional methods. We then introduce community search in complex graphs, focusing on heterogeneous and multilayer networks. Interactive community search is also covered, enabling user feedback and iterative refinement. Finally, we link these techniques to real-world applications and highlight new opportunities from large language models and graph-based retrieval. This tutorial serves as a guide for researchers and practitioners to understand and develop modern community search methods.

Bio. Dr. Longxu Sun is currently a Postdoctoral Research Fellow in the Department of Computer Science at Hong Kong Baptist University. Her primary research interests lie in community search over complex graphs, interactive graph data mining, and AI-augmented graph data management. Her recent work focuses on interactive graph querying algorithms, agentic community mining, and explainable and verifiable community search. Her research has appeared in venues including VLDB, ICDE, TKDE, the SIGMOD Companion, and so on. She has presented a tutorial at PAKDD 2026 and has served as a program committee member for AAAI, WSDM, PAKDD, and IEEE BigData.

Abstract. AI systems increasingly learn from traces of human behavior, including mobility trajectories, health records, online interactions, and social data. While these data enable behavioral modeling, public health insights, and urban mobility analytics, they also create risks of membership and attribute inference, as well as unintended extraction of memorized data. This talk explores how AI can learn from behavioral and social data without revealing protected information across the AI lifecycle. I will highlight our recent work on protecting web-published content before model ingestion, privacy-preserving generation and sharing of mobility and other structured human data, privacy-preserving training, extraction-risk measurement, and machine unlearning. I will conclude with open challenges for trustworthy behavioral AI in generative and agentic systems.

Bio. Li Xiong is the Samuel Candler Dobbs Professor of Computer Science and Biomedical Informatics at Emory University, where she directs the Assured Information Management and Sharing (AIMS) Lab. Her research focuses on trustworthy and privacy-enhancing AI for healthcare, public health, and spatial intelligence. She is an ACM, IEEE, and AAAS Fellow, recognized for her contributions to privacy-preserving and secure data analytics. Her research has been supported by major U.S. government agencies and global industry partners, including Google, IBM, Cisco, AT&T, and Mitsubishi. More details are available at http://www.cs.emory.edu/~lxiong .

Abstract. Smart Space denotes dynamic, adaptive environments enhanced with robotics and AI technologies. Examples include smart homes/offices/cafes. By leveraging and integrating Computer Vision, Natural Language Processing, AIoT, Data Mining, Recommender Systems, and Sympathetic Computing, Smart Space can help improve efficiency, personalization, and user satisfactions with seamless interactions. In this talk, we introduce PolyRAG, a multi-level knowledge QA framework supporting multi-level querying for an indoor robot application system. Building on top of a naive RAG layer, we construct a knowledge pyramid by adding a knowledge graph layer and an ontology schema, so as to obtain a good balance of recall and precision when applied to a specific domain such as coffee robot interactions. We employ cross-layer augmentation techniques for comprehensive knowledge coverage and dynamic updates of the Ontology scheme and instances. To ensure compactness, we utilize cross-layer filtering methods for knowledge condensation in KGs. An experimental coffee robot prototype is constructed, and preliminary empirical studies are conducted to show the effectiveness of our PolyRAG supporting a waterfall model for querying from ontology to KG to chunk-based raw text.

Bio. Qing Li is a Chair Professor and Head of the Department of Computing, the Hong Kong Polytechnic University. He received his B.Eng. from Hunan University (Changsha), and M.Sc. and Ph.D. degrees from the University of Southern California (Los Angeles), all in computer science. His research interests include multi-modal data management, conceptual data modeling, social media, Web services, and e-learning systems. He has authored/co-authored over 500 publications in these areas, with over 74,000 citations and H-index of 111 (source: Google Scholars). He is actively involved in the research community, serving as an Editor-in-Chief of Computer & Education: X Reality (CEXR) by Elsevier; he has also served as an associate editor of IEEE Transactions on Artificial Intelligence (TAI), IEEE Transactions on Cognitive and Developmental Systems (TCDS), IEEE Transactions on Knowledge and Data Engineering (TKDE), ACM Transactions on Internet Technology (TOIT), Data Science and Engineering (DSE), and World Wide Web (WWW) Journal, in addition to being a Conference and Program Chair/Co-Chair of numerous major international conferences. He also sits/sat in the Steering Committees of DASFAA, ACM RecSys, IEEE U-MEDIA, WISE and ICWL. Prof. Li is a Fellow of IEEE, and of International Academy of Artificial Intelligence Sciences (AAIS).

Abstract. The practical deployment of Large Language Models (LLMs) utilizing massive datasets marks a significant milestone in computer science. While this major paradigm shift is characterized by the efficient storage of vast training data and high-performance computing over large-scale computer clusters, it owes much to the achievements deeply rooted in decades of database technology research. Indeed, without the continuous feedback from database research, the progress of AI technology would have been significantly delayed. Conversely, within the database field, the integration of AI into various applications is expanding rapidly. Nevertheless, since database systems inherently require absolute correctness and determinism, applying AI into the core of the database system remains challenging. This talk explores the inherent affinity between AI and database technologies and outlines future perspectives for establishing a mutually beneficial, win-win relationship between both domains.

Bio. Jun Miyazaki received the B.E. degree from the Tokyo Institute of Technology, Tokyo, Japan, in 1992, and the M.S. and Ph.D. degrees from the Japan Advanced Institute of Science and Technology (JAIST), Ishikawa, Japan, in 1994 and 1997, respectively. From 2000 to 2001, he was a visiting scholar in the Department of Computer Science and Engineering at the University of Texas at Arlington. In 2013, he joined the Institute of Science Tokyo (formerly the Tokyo Institute of Technology), Tokyo, Japan, where he is currently a Professor and the Dean of the School of Computing. His research interests include database systems, information retrieval, information recommendation, and high-performance computing. He has been actively involved in numerous international conferences and workshops, serving as a chair or program committee member for events such as IEEE ICDE (2005), VLDB (2020), MDM (2006), SRDS (2014), and DASFAA (2003, 2010, 2024). He also served as Vice President of the Database Society of Japan (DBSJ) from 2018 to 2024. He is a member of ACM, a senior member of IEEE, and a fellow of IEICE and IPSJ.

Abstract. Graph learning has transformed our ability to model complex behavioural and social systems by capturing the relationships, interactions, and dynamics through which individual and collective behaviours emerge. Yet prediction is no longer the final frontier. The rise of large language models (LLMs) and agentic AI raises a more ambitious question: can AI move beyond learning from graphs to understanding social contexts, reasoning about consequences, and acting responsibly? This talk will trace a journey from graph learning to a broader vision of agentic social intelligence. I will discuss how graph representation learning can uncover hidden structures, recurring patterns, and evolving collective dynamics in complex social systems, and also examine critical challenges such as fairness and explainability. I will explore the emerging convergence of graphs, LLMs, and AI agents and conclude by outlining a research agenda for the new frontier of agentic social intelligence.

Bio. Dr. Feng Xia is a Professor of AI at RMIT University, Melbourne, Australia. His research interests include artificial intelligence, graph learning, brain, robotics, and digital health. He is a Fellow of the IEEE. His contributions and leadership have been recognized by prestigious awards. His work is featured in top-tier journals and conferences. Dr. Xia has extensive editorial and organizational experience, having served as an Associate or Guest Editor for over 20 journals and in various Chair roles for more than 30 conferences. He is the Chair of IEEE Task Force on Learning for Graphs. More details are available at http://xia.ai.