Invited Talks

Qingyun Sun 

Title: Graph Machine Learning for the Large Model Era

Abstract:

Graph Machine Learning provides a fundamental paradigm for modeling complex relational data and has achieved remarkable progress over the past decade. This talk revisits several core problems in graph machine learning, with a particular focus on graph structure learning and low-distortion graph representation learning, which aim to uncover latent relational structures and preserve essential structural and semantic information under compact representations. Building upon these foundations, the talk further discusses how graph machine learning is evolving in the era of large models. We present two emerging development roadmaps: graph foundation models, which seek to learn generalizable and transferable representations across diverse graph domains, and Graph Retrieval-Augmented Generation (GraphRAG), which integrates graph-structured knowledge with LLMs to enhance reasoning, retrieval, and generation over complex relational contexts. Together, these directions highlight new opportunities and challenges for advancing graph machine learning toward more scalable, general-purpose, and knowledge-intensive intelligent systems.

Bio: Qingyun Sun is currently an assistant professor at the School of Computer Science and Engineering at Beihang University, China. She has published over 60 papers on graph learning and social network mining on TPAMI, TKDE, NeurIPS, ICML, ICLR, WWW, etc. She was the recipient of IEEE IWQoS 2022 Best Paper Award, CIKM 2022 Best Paper Honorable Mention Award, ICDM 2021 Best Paper Candidate, and the most influential WWW papers. She serves as the area chair and program committee member of NeurIPS, ICLR, ICML, AAAI, IJCAI, WWW, KDD, ICDM, etc. She have organized and hosted challenges at IJCAI 2025 and IEEE BigData 2024, a workshop at CIKM 2025, and delivered a Tutorial at IJCAI 2025.

Xiao Li

Title: Graph Analysis: From Molecules to Time Series

Abstract: 

Network analysis offers a powerful, unifying lens for modeling complex systems across scientific domains. This talk will chart a journey from the static networks of bioinformatics to the dynamic graphs of multivariate time series, showcasing the versatility of graph-structured thinking.

We begin by exploring how heterogeneous information networks integrate rich, multi-typed biological data—such as protein interactions, genomic sequences, and drug-target relationships. We demonstrate how this framework tackles two critical predictive tasks: identifying disease-associated genes and discovering novel drug-target interactions. Here, the graph itself provides the prior knowledge that effectively narrows the search space for costly experimental validation.

We then shift from what is connected to how connections evolve. Modern sensor systems generate multivariate time series where the relationships between sensors are dynamic and latent. By reframing these sequences as spatial-temporal graphs, we move beyond traditional sequential models. This approach explicitly captures both intra-sensor dependencies and the evolving inter-sensor interactions, leading to more powerful forecasting and pattern recognition.

Together, this talk illustrates how the core paradigm of network analysis—modeling entities as nodes and relationships as edges—provides a remarkably versatile toolkit. It bridges domains from decoding the static complexity of biological systems to mastering the dynamic, hidden patterns in real-world sensor data for real-world applications.

Bio: Xiaoli is currently a Full Professor and Head of the Information Systems Technology and Design Pillar at Singapore University of Technology and Design (SUTD). He previously led A*STAR’s Machine Intellection Department, where he built and directed Singapore’s largest AI and data science research group. He is also an Adjunct Full Professor at Nanyang Technological University, and a Fellow of both IEEE and AAIA.

His research spans AI, data mining, machine learning, and bioinformatics, and has produced more than 400 peer-reviewed publications with over 30,000 citations, an h-index of 90, and more than ten best paper awards. He serves as Editor-in-Chief of the Annual Review of Artificial Intelligence and as an Associate Editor for leading journals such as IEEE Transactions on Artificial Intelligence and Knowledge and Information Systems. He has also played key leadership roles as conference chair or area chair at premier venues including AAAI, IJCAI, ICLR, NeurIPS, KDD, and ICDM.

Beyond academia, Xiaoli brings extensive industry engagement experience, having established and led multiple joint labs and spearheaded more than ten major R&D collaborations with global partners in aerospace, telecommunications, insurance, and professional services.

His contributions have earned him international recognition as one of the world’s top 2% scientists in AI (Stanford University) and as a Clarivate Highly Cited Researcher.

FANG Yuan

Title: Semantic-Structural Integration in Text-Attributed Graphs

Abstract: 

Graphs enriched with textual information—known as text-attributed graphs (TAGs)—are increasingly common in domains such as scientific research, online platforms, and knowledge networks. They combine two complementary perspectives: the semantic content conveyed by text and the structural patterns encoded in graph connections. Yet, unifying these two sources of information in a seamless and effective manner remains a fundamental challenge. This talk presents recent advances in bridging semantics and structure within TAGs. First, graph-grounded pre-training and prompting enhance low-resource text classification by jointly training a graph and a text model through node-text alignment. Second, in molecular analysis, an unsupervised fine-grained alignment approach links molecular substructures with corresponding textual segments, enabling generalization to unseen molecules and supporting complex scientific applications. Third, a quantization-based framework transforms graph information into discrete tokens, ensuring compatibility with large language models and supporting true zero-shot transfer across datasets. Collectively, these developments chart a path toward a new generation of methods that integrate structural and semantic information, offering greater adaptability, interpretability, and impact across diverse domains.

Bio: FANG Yuan is currently an Associate Professor (tenured) at the School of Computing and Information Systems, Singapore Management University (SMU). Prior to joining SMU, he was a data scientist at DBS Bank, and a research scientist at A*STAR. He obtained a PhD Degree in Computer Science from the University of Illinois at Urbana-Champaign in 2014, and Bachelor of Computing with First Class Honors from National University of Singapore in 2009. His general research interest lies in the broad areas of data mining, machine learning and artificial intelligence. More specifically, he is working on graph machine learning, as well as its applications in Web and social media mining, recommendation systems, and bioinformatics. He has published over 100 paper in leading conferences and journals, with his research featured in VLDB’13 Best Papers collection and the Most Influential WWW’23 paper (Paper Digest). He has also been recognized as a Stanford’s World Top 2% Scientist (Single Year, 2024--2025). He actively contributes to the research community through editorial and organizational roles. He has served as a Young Associate Editor for Frontiers of Computer Science, and Area Chairs/Senior PC Members of leading conferences including KDD, WWW, NeurIPS, ICLR, etc. He has also served as the Secretary of the Singapore Chapter of ACM SIGKDD, and served on the Organizing Committee of KDD’21, WSDM’23 and WWW’24.

Zhiguang Cao

Title: Learning to Solve Vehicle Routing Problems on Graphs

Abstract: 

Vehicle routing problems (VRPs) are fundamental combinatorial optimization problems in operations research, naturally represented as graphs and traditionally solved using carefully designed heuristics and handcrafted rules. In recent years, there has been increasing interest in learning-based approaches that leverage deep neural networks, and more recently, large language models (LLMs), to automatically discover effective solution strategies. In this talk, I will first review representative neural approaches for solving individual VRP variants. I will then discuss recent progress toward more generalizable models that aim to transfer knowledge across multiple VRP variants, moving in the direction of foundation models for routing problems. Finally, I will outline key open challenges and share my perspectives.

Bio:  Dr. Zhiguang Cao is an Assistant Professor at the School of Computing and Information Systems, Singapore Management University (SMU). He received his Ph.D. from Nanyang Technological University, Singapore. His research centers on AI for Optimization (AI4Opt), where deep learning techniques (including LLM) are applied to solve classical combinatorial optimization problems such as the vehicle routing, job-shop scheduling, and bin packing. Dr. Cao has published 30+ papers at ICML, NeurIPS and ICLR, where he also serves regularly as an Area Chair. More information about Dr. Cao can be found here: https://zhiguangcaosg.github.io/

Liu Fayao

Title: Beyond the Grid: Scaling Structural Reasoning from Monocular Depth to 4D Scenes and CAD Synthesis

Abstract: 

Deep learning has traditionally treated visual data as a simple grid of pixels. However, the physical world is governed by complex structures—from the geometric relationships between surfaces to the logical "recipe" used to manufacture an object. To achieve true physical reasoning, AI must look "beyond the grid" and learn to represent these underlying structures explicitly.

In this talk, I will share our work on scaling structural reasoning across three levels of increasing complexity. We begin with monocular depth estimation, where we use Deep Convolutional Neural Fields to model relational dependencies between pixels. We then move to 4D spatio-temporal reasoning, introducing a progressive grouping method that discovers structure in massive, weakly labelled outdoor point clouds. Finally, we explore the peak of structural complexity with CADCrafter, a model that generates parametric CAD sequences from unconstrained real-world images.

Bio: Dr. Fayao Liu is a Research Scientist at the Institute for Infocomm Research (I²R), A*STAR, Singapore. Her research lies at the intersection of artificial intelligence and computer vision, with a focus on 3D vision, generative models, and multimodal representation learning. Her work aims to enable machines to perceive, reason about, and generate structured 3D worlds from visual and multimodal data. She serves as an area chair and associate editor for leading computer vision conferences and journals and was named an honoree of Singapore’s 100 Women in Tech in 2023.

Xavier Bresson

Title: Integrating Large Language Models and Graph Neural Networks

Abstract: 

Pre-trained language models on large-scale datasets have revolutionized text-based applications, enabling new capabilities in natural language processing. When documents are connected, they form a text-attributed graph (TAG), like the Internet, Wikipedia, social networks, scientific literature networks, biological networks, scene graphs, and knowledge graphs. Key applications for TAGs include recommendation (web), classification (node, link, graph), text- and visual-based reasoning, and retrieval-augmented generation (RAG). In this talk, I will introduce two approaches that integrate Large Language Models (LLMs) with Graph Neural Networks (GNNs). The first method demonstrates how LLMs’ reasoning capabilities can enhance TAG node features. The second approach introduces a pioneering technique called GraphRAG, which grounds LLM responses in a relevant sub-graph structure. This scalable technique regularizes the language model, significantly reducing incorrect responses, a.k.a. hallucinations.

Bio: Xavier Bresson is an Associate Professor in the Department of Computer Science at the National University of Singapore (NUS). His research centers on graph deep learning, bridging graph theory and neural networks to model complex relational data. In 2017, he received a USD 2M NRF Fellowship to advance this area, alongside additional research grants in the U.S. and Hong Kong. He has co-authored one of the field’s most influential papers (among the top-cited at NeurIPS) and has played a key role in shaping modern graph learning methods. He regularly organizes and contributes to major conferences, workshops, and tutorials (including Learning on Graphs and events at NeurIPS, CVPR, MLSys, KDD, AAAI, ICML, and ICLR).

TAY, Wee Peng

Title: Continuous Dynamics in Graph Neural Networks

Abstract:

Graph neural networks (GNNs) have been successfully applied across a wide range of domains, and recent years have witnessed a rapid proliferation of GNN architectures. In this talk, I focus on a class of GNN models grounded in continuous-time dynamics and demonstrate several advantages of this formulation. By leveraging tools from dynamical systems theory, these models admit principled interpretability and enable rigorous theoretical analysis. In particular, they are equipped with provable stability guarantees, which in turn yield robustness guarantees against adversarial perturbations. Moreover, certain architectures can provably mitigate over-smoothing, support long-range interactions, and facilitate contrastive representation learning without requiring data augmentations or negative samples. I will present the theoretical foundations underlying these phenomena and introduce several physics-inspired GNN architectures, including Hamiltonian graph diffusion and fractional-order diffusion models. Finally, I will illustrate their effectiveness through selected applications.

Bio: Wee Peng Tay received the B.S. degree in Electrical Engineering and Mathematics, and the M.S. degree in Electrical Engineering from Stanford University in 2002. He received the Ph.D. degree in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology in 2008. He is currently a Professor of Signal and Information Processing in the School of Electrical and Electronic Engineering at the Nanyang Technological University, Singapore. Dr. Tay received the Tan Chin Tuan Exchange Fellowship in 2015, the IEEE Signal Processing Society Young Author Best Paper Award with his student in 2016, and several best conference paper awards. He has served as an Associate Editor or Editor for several journals, including the IEEE Transactions on Signal Processing, IEEE Transactions on Signal and Information Processing over Networks, and the IEEE Transactions on Wireless Communications. He is currently an Associate Editor for the IEEE Internet of Things Journal, and a Subject Editor for Signal Processing, Elsevier. is research interests include signal and information processing over graphs, and robust machine learning. 

Mengting Wan 

Title: From Interaction Graphs to Collaboration Intelligence: Learning from Higher‑Order, Semantic-Rich Interactions in the Wild

Abstract:

Collaboration is among the most structurally complex and semantically rich forms of human interaction. Unlike engagement-driven consumer social networks, collaboration in productivity settings is one of the most economically consequential forms of social interaction: outcomes are collective, stakes are high, and failures propagate across teams and organizations. Participants operate under evolving roles, expectations, and social norms, while models need to infer latent intent, decide when and how to act, and coordinate interventions across time.

We argue that recent shifts toward digital and hybrid work have made the topology of collaboration increasingly observable. Interaction traces from shared workspaces (e.g., group chat, co-editing) naturally induce higher-order and temporal graph structures that capture group dynamics beyond static pairwise ties. Large language models (LLMs) further unlock the semantic layer embedded in these noisy interaction graphs. By inferring intent, norms, friction, and social context from unstructured traces, LLMs enable semantically grounded representations of collaboration and transform these signals into operational reward signals through in-situ user feedback.

Together, these developments point toward an emerging notion of collaboration intelligence: systems that can reason over higher-order, semantic-rich interaction graphs, act with appropriate timing and permission, and contribute toward shared outcomes. Getting there will require that we, as a community, move beyond standard relational abstractions and elevate semantic-rich network interactions as a first-class route to collective intelligence.

Bio: Mengting Wan is a Principal Research Scientist in Microsoft’s Office of Applied Research, where she works on advancing AI systems for complex, real‑world collaborative behaviors. Her research spans LLM social reasoning, agentic systems, and reinforcement learning, with a recent focus on training multiparty collaborative agents and simulating realistic multiparty user interactions. Previously, Mengting conducted extensive research in recommender systems, graph machine learning, and network science. She publishes extensively at top machine learning venues and has received multiple honors, including the KDD 2022 Best Research Paper Award. Mengting holds a Ph.D. in Computer Science from UC San Diego, where she was a Microsoft PhD Fellow from 2017 to 2019.