Graph learning · Scientific & biomedical AI

I develop domain-knowledge-adaptive AI: models that combine the relationships in data with scientific knowledge to make predictions under noise, limited observations, and computational constraints.

My methodological foundation is robust and efficient graph learning. My current research extends this foundation to molecular property prediction, biological dynamics, and longitudinal PET imaging, with a focus on forecasting regional tau changes in Alzheimer’s disease.

I am a Postdoctoral Associate at Yale University, working with Georges El Fakhri, Jingsong Ouyang and Xiaofeng Liu. I received my Ph.D. in Computer Science from the University of Hong Kong, advised by Siu-Ming Yiu, and was a visiting postdoctoral researcher at the University of Cambridge with Pietro Liò.

My central research question: How can AI combine data and scientific knowledge to better predict real-world outcomes?

Two connected research lines

RESEARCH LINE 01

Graph Neural Networks for Robust and Efficient Spatiotemporal Learning

Problem. Real systems contain interacting entities that change over time. Their measurements are incomplete, their relationships can be noisy, and large graph models can be costly to run.

Approach. I develop adaptive graph augmentation and contrastive learning to extract useful relational signals from imperfect data. I also study knowledge distillation and efficient temporal architectures to preserve predictive information with lower computational cost.

Applications. Urban sensing, traffic forecasting, trajectory analysis, and recommendation provide settings for studying robustness, heterogeneous relationships, and temporal prediction at scale.

Representative work: AutoST · WWW 2023; STAG · ICML 2023; LightST · AAAI 2025; AutoHFormer · ICDE 2026.

RESEARCH LINE 02

Knowledge-Guided AI for Biological Discovery and Precision Medicine

Problem. Biological and medical datasets are often small, heterogeneous, and expensive to acquire. Predictions must account for scientific structure and remain useful as measurement conditions change.

Approach. I investigate models that adapt quantum, geometric, biological, and anatomical priors to the task. My current work combines graph learning, generative modeling, and multimodal representations to study molecular properties, cellular dynamics, and longitudinal imaging.

Current focus. Forecasting future regional tau PET, incorporating connectivity and scan quality, and calibrating probabilistic forecasts across cohorts. Integrating imaging with patient genomics is a future direction.

Representative work: molecular learning survey and benchmarks · JCTC 2026; OG-QIMP, BioDM, and longitudinal tau PET studies · ongoing / under review.

Why these methods address the problem

Each part of my approach targets a specific failure mode in learning from complex systems.

Represent interactions
Graphs make dependencies between entities explicit. Learning across related entities captures information that isolated observations miss, while heterogeneous relations preserve different kinds of connection.
Adapt to imperfect data
Learned graph augmentation and contrastive objectives provide supervision from the data itself. They help models learn useful representations when observations or graph connections are noisy and incomplete, as studied in AutoST and STAG.
Use scientific structure
Scientific priors provide structure when observations alone are insufficient. My biomedical agenda asks how to adapt those priors to each task and cohort, and how to evaluate when they help or fail.
Make prediction practical
LightST transfers spatial and temporal knowledge into a lightweight predictor. AutoHFormer organizes temporal computation hierarchically. Both address the cost of learning and predicting over complex sequences.

An independent research agenda

My contribution brings together method development in graph learning, efficient temporal modeling, and scientific problem formulation. This creates a common research agenda across computer science and biomedicine: learning adaptive representations of interacting systems, then testing their robustness, efficiency, and transfer across settings.

I aim to build a research group that advances this agenda through reusable algorithms and rigorous benchmarks. In biomedicine, my next steps are to evaluate longitudinal imaging forecasts across cohorts and investigate how imaging and genomic information can support more individualized models of disease progression.

Selected publications

Selected contributions are grouped by research line; manuscripts under review are listed separately. Full publication list on Google Scholar.

Graph and temporal learning

  1. Efficient Prompt Learning for Traffic PredictionRegina Zhang et al. · The VLDB Journal (VLDBJ), 2026 · First author
  2. AutoHFormer: Efficient Hierarchical Autoregressive Transformer for Time Series PredictionRegina Zhang et al. · International Conference on Data Engineering (ICDE), 2026 · First authorHierarchical temporal modeling and causal windowed attention address the computational cost of long-horizon forecasting.
  3. M²Rec: Multi-scale Mamba for Sequential RecommendationRegina Zhang et al. · IEEE Transactions on Knowledge and Data Engineering (TKDE), 2026 · First author
  4. HMamba: Hyperbolic Mamba for Sequential RecommendationRegina Zhang et al. · ACM Transactions on Information Systems (TOIS), 2026 · First author
  5. Efficient Traffic Prediction Through Spatio-Temporal DistillationRegina Zhang et al. · AAAI Conference on Artificial Intelligence, 2025 · Oral · First authorLightST transfers graph knowledge to a lightweight student. The paper reports 5–40× faster prediction than the compared spatiotemporal GNNs in its traffic benchmarks, with superior accuracy.
  6. HGAurban: Heterogeneous Graph Autoencoding for Urban Spatial-Temporal LearningRegina Zhang et al. · Conference on Information and Knowledge Management (CIKM), 2025 · First author
  7. Graph Augmentation for RecommendationRegina Zhang et al. · International Conference on Data Engineering (ICDE), 2024 · First author
  8. Automated Spatio-Temporal Graph Contrastive LearningRegina Zhang et al. · The Web Conference (WWW), 2023 · First authorAutoST learns graph augmentations over multiple views to address noise, missing data, and heterogeneous spatial and temporal patterns.
  9. Spatial-Temporal Graph Learning with Adversarial Contrastive AdaptationRegina Zhang et al. · International Conference on Machine Learning (ICML), 2023 · First authorSTAG adapts contrastive supervision and hard-sample selection to improve graph representations under imperfect observations.
  10. Online Anomalous Subtrajectory Detection with Road Network Enhanced Reinforcement LearningRegina Zhang et al. · International Conference on Data Engineering (ICDE), 2023 · First author

Biological and biomedical AI

  1. A Systematic Survey and Benchmarks of Deep Learning for Molecular Property Prediction in the Foundation Model EraZongru Li et al. · Journal of Chemical Theory and Computation (JCTC), 2026 · Corresponding author · Cover paperA survey and benchmarking contribution to molecular learning in the foundation model era.
  2. An Open Multi-Center Whole-Body FDG PET/CT Foundation Model for Tumor SegmentationXiaofeng Liu, Regina Zhang et al. · CVPR 2026, Multimodal Foundation Models for Biomedicine workshop
  3. Longitudinal Medical Visual Question Answering Guided by Vision Foundation Models for Consistent AttentionJ. Wu, Regina Zhang et al. · CVPR 2026, PHAROS AI Factory for Medical Imaging and Healthcare workshop
Further work on foundation models, time series, and evaluation

Teaching & professional service

Teaching

Teaching Assistant, FITE7405: Techniques in Computational Finance, University of Hong Kong, four offerings during 2021–2024.

My teaching interests include graph machine learning, time-series analysis, and AI for scientific and biomedical applications.

Service

Reviewer for venues including TKDE, ICLR, ICML, NeurIPS, AAAI, KDD, and The Web Conference.

Time Series Session Chair, SIAM International Conference on Data Mining (SDM), 2022.

Selected talks & honors

Invited talks

  • AAAI workshop on neural-symbolic NLP and knowledge graph reasoning · Keynote, 2026
  • University of Cambridge · 2025
  • Yale University · 2024
  • Cornell University · 2024
  • Westlake University · 2023

Honors

  • Baidu Research Fellowship · One of 40 recipients worldwide, 2023
  • Student Travel Award · WWW 2023
  • Postgraduate Scholarship · HKU, 2020–2024

Research opportunities & collaboration

I welcome inquiries from prospective research interns interested in graph learning, time-series modeling, molecular AI, or biomedical imaging. Strong programming skills and an interest in rigorous experimentation are especially valuable.

Please email your CV, GitHub profile, and a short description of your interests to reginazhang955@gmail.com. I also welcome collaborations on robust, efficient, and knowledge-guided AI.