HKUST(GZ) × UM Data Science Workshop

Workshop Schedule

Time
14:30-14:40	Workshop Opening
14:40-15:00	Optimizing Scalable LLM Inference: System-Level Strategies for Proactive KV Cache Management Prof. Lei Chen, Hong Kong University of Science and Technology (Guangzhou)
15:00-15:20	Graph Learning for Recommender Systems Prof. Zhiguo Gong, University of Macau
15:20-15:40	Any-File to Any-Form: The Coming Revolution in Structured AI Prof. Nan Tang, Hong Kong University of Science and Technology (Guangzhou)
15:40-16:00	How AI Techniques Could Reform Databases: Optimizing Data Systems Through Learning Prof. Leong Hou U, University of Macau
16:00-16:20	Coffee Break
16:20-16:40	Large Language Model Empowering AI Agents: Tackling Core Challenges and Advancing Real-World Applications Prof. Jing Tang, Hong Kong University of Science and Technology (Guangzhou)
16:40-17:00	Animating the Crowd: Building Mobility Digital Twins for Smart Campuses Prof. Dingqi Yang, University of Macau
17:00-17:20	Empowering LLMs: How Graph Reasoning Drives Generalized Thinking Prof. Jia Li, Hong Kong University of Science and Technology (Guangzhou)
17:20-17:40	Understanding Distribution Learning of Diffusion Models via Subspace Clustering Prof. Peng Wang, University of Macau

Speakers from HKUST(GZ)

Prof. Lei Chen

Chair Professor of HKUST(GZ)

Optimizing Scalable LLM Inference: System-Level Strategies for Proactive KV Cache Management

Abstract: As large language models (LLMs) increasingly underpin mission-critical applications across industries, optimizing their inference efficiency has emerged as a critical priority. Central to this optimization is the effective management of the Key-Value (KV) cache, a memory-intensive component that stores intermediate computations to accelerate autoregressive token generation.
In this talk, we examine recent advancements in system-level KV cache management, emphasizing novel approaches to proactive scheduling that dynamically allocate computational and memory resources. We evaluate techniques optimized for diverse operational contexts—spanning offline batch processing to real-time online serving—and discuss architectural optimizations for single-instance execution as well as coordination strategies for concurrent multi-instance deployments. Finally, we outline promising research directions to address scalability challenges in multi-instance inference. These advancements are crucial for enabling scalable enterprise solutions as LLMs expand into latency-sensitive, high-throughput industrial applications.

Lei Chen is a Chair Professor in Data Science and Analytics at HKUST (GZ), a Fellow of ACM and IEEE. Currently, he serves as the Dean of the Information Hub and the Director of the Big Data Institute at HKUST (GZ). Prof. Chen’s research spans several areas, including Data-driven AI, Big Data Analytics, the Metaverse, knowledge graphs, blockchain technology, data privacy, crowdsourcing, and spatial and temporal databases, as well as probabilistic databases. He earned his Ph.D. in Computer Science from the University of Waterloo, Canada.
Prof. Chen has received several prestigious awards, including the SIGMOD Test-of-Time Award in 2015 and the Best Research Paper Award at VLDB 2022. His team’s system also won the Excellent Demonstration Award at VLDB 2014. He served as the Program Committee Co-chair for VLDB 2019 and currently holds the position of Editor-in-Chief for IEEE Transactions on Data and Knowledge Engineering. In addition, he was the General Co-Chair of VLDB 2024 and will serve as the General Co-Chair of IJCAI China 2025.

Prof. Nan Tang

Associate Professor of HKUST(GZ)

Any-File to Any-Form: The Coming Revolution in Structured AI

Abstract: Most of the world’s knowledge remains locked in unstructured documents—reports, PDFs, slides, and images. This talk envisions a future where AI can turn any file into any structured form—databases, graphs, timelines, or knowledge maps—and transform these structures into meaningful insights. I will introduce emerging techniques that blend large language models with constraint-aware verification to extract, transform, and reason over structured data reliably. This new paradigm of Structured AI bridges content understanding and data analytics, redefining how we move from raw documents to trustworthy, explainable, and interactive data intelligence.

Dr. Nan Tang is currently an Associate Professor at the Hong Kong University of Science and Technology (Guangzhou) (2023-). He received his Ph.D. in Systems Engineering and Engineering Management from The Chinese University of Hong Kong in December 2007. He has worked at CWI (2008-2009), University of Edinburgh (2010-2011), QCRI (2011-2023). He has served extensively in the academic community as reviewer, area chair, and conference chair for top-tier venues, including VLDB 2026 & 2025, ICDE 2026 & 2024, KDD 2025 (Datasets and Benchmarks), CIKM 2025 & 2024, and as Exhibition Chair of SIGMOD 2021 and Demo Chair of DASFAA 2019. His work has been recognized with numerous prestigious awards, including the SIGMOD 2024 Research Highlight Awards, Best Paper Awards at SIGMOD 2023, ICDE (2018, 2012, 2009), VLDB (2023, 2015, 2010), the ACM SIGMOD 2020 Reproducibility Award, and the VLDB 2021 Distinguished Reviewer Award.

Prof. Jing Tang

Assistant Professor of HKUST(GZ)

Large Language Model Empowering AI Agents: Tackling Core Challenges and Advancing Real-World Applications

Abstract: In the new era of generative AI, LLM-driven agents have delivered monumental success by achieving human-level fluency in language, reasoning, and code generation. However, their pervasive application faces significant challenges, including the risk of hallucinations, data privacy concerns from massive training sets, and the need for reliable controllability. To mitigate these issues, we offer a comprehensive exploration of key technologies and integrated solutions across the “Data-Model-Interaction” stack. For data governance, we employ Retrieval-Augmented Generation (RAG) and synthetic data to expand knowledge bases and to incorporate multimodal information, significantly mitigating factual errors. For model optimization, we achieve domain specialization through Parameter-Efficient Fine-Tuning (PEFT), deploy model distillation to compress massive and multi-billion parameter LLMs, and boost decision-making robustness via Reinforcement Learning (RL) mechanisms. For interaction control, we adopt “native-speaking” Chain-of-Thought (CoT) prompting and GUI agents to enable precise and controllable behavior. Ultimately, these advancements aim to drive the successful deployment of AI agents in quantitative trading and intelligent companionship.

Jing Tang is an Assistant Professor in the Data Science and Analytics Thrust, the Program Director of MSc DCAI, and the Director of Center for Blockchain Technology and Digital Media at The Hong Kong University of Science and Technology (Guangzhou). He is also an Assistant Professor in the Division of Emerging Interdisciplinary Areas at The Hong Kong University of Science and Technology. He received his Ph.D. degree from the Nanyang Technological University (2013–2017) and his B.Eng. degree from the University of Science and Technology of China (2008–2012). Prior to joining HKUST(GZ), he was a Research Assistant Professor at National University of Singapore (2018–2021). His research focuses on big data management and analytics, social network and graph analysis, machine learning, large language models, and blockchains. He has published more than 70 high-quality research papers, mainly in SIGMOD, VLDB, ICDE, ICML, NeurIPS, ICLR, KDD, TODS, VLDBJ, and TKDE. He has been honored with CES 2025 Innovation Award (AI Category), Gold Medal at the 2025 Geneva International Exhibition of Inventions, Best Paper Award from the IEEE ICNP in 2014, the Best-in-Session Presentation Award from the IEEE INFOCOM in 2018, and finalist for the Singapore NRF Fellowship for AI in 2019.

Prof. Jia Li

Assistant Professor of HKUST(GZ)

Empowering LLMs: How Graph Reasoning Drives Generalized Thinking

Abstract: In recent years, large language models (LLMs) have been widely used to enhance the generalization of graph models, both across graph tasks and domains—a trend known as “LLM for Graph.” Recently, especially with the advent of models like Deepseek R1, research focus has shifted toward “Graph for LLM”: leveraging graph reasoning tasks to fundamentally improve the general reasoning capabilities of LLMs. Graph reasoning, with its inherent structural complexity and multi-step logic, provides an ideal testbed for advancing LLMs’ abilities in mathematical, logical, and commonsense inference. This talk will explore why graph reasoning is a key scenario for boosting LLMs’ general reasoning skills, and discuss the latest progress and future directions in this emerging field.

Dr. Jia Li is an Assistant Professor in Data Science and Analytics Thrust of HKUST (Guangzhou). He also serves as the Co-Director of HKUST-GZ Create Link Joint Lab. He received a Ph.D. degree at The Chinese University of Hong Kong in 2021. During the Ph.D. study, he worked as a research intern at Google AI (Mountain view) and Tencent AI Lab. Before the Ph.D. study, he worked as a full-time data mining engineer in Tencent from 2014 to 2017, where he was responsible for WeChat and WeBank fraud prevention. He received several awards including KDD Best Paper Award in 2023 and Tencent Rhino Bird Excellent Research Award in 2022.

Speakers from UM

Prof. Zhiguo Gong

Professor of University of Macau

Graph Learning for Recommender Systems

Abstract: Despite years of development, recommender algorithms continue to attract significant attention from both academic and industry communities, driven by ongoing practical demands. In recent years, Graph Neural Network (GNN) techniques have been increasingly integrated into recommender system models, offering new capabilities and insights. In this talk, I will provide a concise overview of the evolution of graph-based learning approaches in recommender systems, along with some findings from our recent research in this area.

Professor Zhiguo Gong is a professor in the StateKey Lab of Internet of Things for Smart City (IOTSC) and the Department of Computer and Information Science (CIS) at the University of Macau. He currently serves as the Associate Dean of the Faculty of Science and Technology. Professor Gong earned his Ph.D. from the Chinese Academy of Sciences in 1998. His research interests include data mining and machine learning. Professor Gong has held prominent roles in academic events, serving as General Co-Chair of IEEE/WIC/ACM WI/IAT 2012, WAIM 2014, and ApWeb-WAIM 2018, as well as Local Chair for ICDE 2019, IJCAI 2019, and IJCAI 2023. He was Program Committee Chair for PAKDD 2019 and IEEE ICBK 2021.

Prof. Leong Hou U

Associate Professor of University of Macau

How AI Techniques Could Reform Databases: Optimizing Data Systems Through Learning

Abstract: The convergence of artificial intelligence and database technologies is redefining modern data management. Learning-driven methods enhance scalability, expressivity, and adaptability across data systems. Reinforcement learning automates query and index optimization, while advanced graph neural networks and structural embeddings improve relational representation and retrieval accuracy. Vector search and diversity-based exploration enable adaptive, efficient query handling, and out-of-distribution detection strengthens reliability in real-world deployment. Finally, geometric and metamodal latent-space modeling fosters consistent, semantically aligned multi-modal data integration. Together, these advances move databases from static storage engines toward intelligent, self-optimizing platforms that can understand, organize, and generate knowledge at scale.

Leong Hou U, Director of the Centre for Data Science at the University of Macau, leads a research team specializing in AI and data engineering with a focus on spatiotemporal data, graph learning algorithms, reinforcement learning, traffic optimization, and intelligent systems. With over 80 publications in top-tier conferences such as SIGMOD, VLDB, and NeurIPS, his team has contributed to several national and regional key R&D projects, including COVID-19 data analytics and smart city infrastructure. Their work combines high-performance computing with real-world urban applications, earning a Second-Class Science and Technology Invention Award in Macau and extensive media recognition. Prof. U is also actively involved in international academic leadership and local governance advisory roles.

Prof. Dingqi Yang

Associate Professor of University of Macau

Animating the Crowd: Building Mobility Digital Twins for Smart Campuses

Abstract: Understanding crowd mobility is critical for many smart city applications, such as urban planning, resource allocation, and emergency management. In this talk, I will share our recent project on building a mobility digital twin for smart campuses. I will first introduce our mobility data sensing and extraction pipeline from comprehensive but noisy WiFi connection logs on a university campus, followed by our recent research efforts on designing not only the predictive but also simulative mobility models for both collective crowd flows and individual trajectories. Meanwhile, by revisiting the existing benchmarking protocols for mobility digital twins, a task-based evaluation protocol is introduced for assessing the ultimate utility of synthetic human mobility data. Finally, a prototype system is also built supporting both predictive and simulative analyses of on-campus crowd mobility.

Dingqi Yang is an Associate Professor with the State Key Laboratory of Internet of Things for Smart City and the Department of Computer and Information Science, University of Macau. He received his Ph.D. degree in Computer Science from Pierre and Marie Curie University and Institut Mines-Télécom in France, where he won both the CNRS SAMOVAR Doctorate Award and the Institut Mines-Télécom Press Mention in 2015. Before joining the University of Macau, he worked as a senior researcher at the University of Fribourg in Switzerland from 2015 to 2020. His research interests include spatiotemporal data analytics, knowledge graphs, ubiquitous computing, and smart city applications. More details can be found on his homepage.
(https://www.fst.um.edu.mo/personal/yangdingqi/)

Prof. Peng Wang

Assistant Professor of University of Macau

Understanding Distribution Learning of Diffusion Models via Subspace Clustering

Abstract: Diffusion models have recently emerged as a powerful new family of generative models. Remarkably, they are capable of learning image distributions to generate high-quality images that are distinct from the training data, even when trained on a finite dataset. However, worst-case analyses have shown that learning arbitrary distributions via diffusion models may require exponentially many samples in terms of the ambient data dimension, suggesting a gap between theory and practice. In this work, we close this gap by studying how diffusion models learn low-dimensional distributions, motivated by the observation that real-world image datasets often exhibit intrinsic low-dimensional structures. Specifically, we focus on learning a mixture of low-rank Gaussians, where the low-rank covariance captures the underlying low dimensionality of the data and the latent Gaussian structure ensures theoretical tractability. Theoretically, under an appropriate network parameterization, we rigorously show that optimizing the training loss of diffusion models is equivalent to solving the canonical subspace clustering problem over the training samples. Based on this equivalence, we further demonstrate that the minimal number of samples required to learn the distribution scales linearly with the intrinsic dimension under our data and model assumptions. This insight sheds light on why diffusion models can break the curse of dimensionality and exhibit phase transitions in learning distributions. Empirically, we validate our results with various experiments on synthetic and real-world image datasets. Moreover, based on our theoretical findings, we establish a correspondence between the subspace bases and semantic attributes of image data, providing a foundation for controllable image editing.

Peng Wang is an assistant professor at the Department of Computer and Information Science at the University of Macau. Before this, he is a postdoc research fellow advised by Professors Laura Balzano and Qing Qu at the University of Michigan. He got the Ph.D. degree under the supervision of Professor Anthony Man-Cho So at The Chinese University of Hong Kong. His research interest lies in the intersection of optimization, machine learning, and artificial intelligence.