Research
My research spans the foundations of machine learning and its applications. Four connected directions run through my work, many of them linked by a common thread: learning from structure — in graphs, time series, language, and health data.
Research directions
Graph Neural Networks
Deep graph learning analyzes data with graph structure, such as social networks, drug–target interaction networks, and molecules. My group works on making graph neural networks more powerful, efficient, and adaptable: scalability, dynamic graphs, graph generation, learning with geometry and topology, graph coarsening, graph pretraining, and interpretable graph models.
Representative work
- FastGCN ICLR 2018
- EvolveGCN AAAI 2020
- Constrained GraphVAE NeurIPS 2018
- Illinois Graph Benchmark KDD 2023
- Persistent homology on graphs ICML 2021, NeurIPS 2022, JMLR 2025
- Graph Concept Bottleneck Models TMLR 2026
Time Series Learning
Time series are everywhere in healthcare and science, but deep models for them are often hard to interpret and hard to transfer. We build neuro-symbolic and shape-based models that are interpretable by design, time-series foundation models that generalize across domains, and causality-aware methods for anomaly detection and prediction.
Representative work
- Abstracted Shapes as Tokens NeurIPS 2024
- Interpretability Gated Networks ICLR 2025
- Weighted Clock Logic Point Process ICLR 2023
- Neuro-symbolic time series classification ICDM 2022
- CAAD: causality-aware anomaly detection KDD 2026
Language & Intelligence
I have worked on natural language processing since my early career. Earlier work focused on document representations and summarization; more recently the focus has shifted to the trustworthiness and transparency of large language models — how they use structure, how robust they are, and how to make them efficient — as well as AI for code and medical NLP.
Representative work
- Resting Neurons, Active Insights ICML 2026
- When Structure Doesn't Help: LLMs and text-attributed graphs LoG 2025
- Attention abnormality in Trojaned BERTs NAACL 2022
- Timeline summarization via optimal transport EMNLP 2021
- LLM-generated code detection AAAI 2025
AI for Healthcare
Many graph learning problems come from biomedicine, and I also work directly on computational healthcare. With collaborators, we develop graph, neuro-symbolic, and time-series models for patient phenotyping and risk prediction, readmission prediction, medication recommendation, drug–drug interactions, EHR question answering, and machine-learning-guided wound healing.
Representative work
- GAMENet: medication recommendation AAAI 2019
- Pre-trained graph-augmented transformers (G-BERT) IJCAI 2019
- Multi-view graph auto-encoders for drug similarity IJCAI 2018
- HypEHR: hyperbolic EHR question answering ACL 2026 Findings
- Battery-free wireless wound monitor Adv. Healthc. Mater. 2023
A cross-section of my work
Selected publications
- 2026ACL FindingsHypEHR: Hyperbolic Modeling of Electronic Health Records for Efficient Question Answering
- 2026TMLRGraph Concept Bottleneck Models
- 2025KDDUnderstanding and Tackling Over-Dilution in Graph Neural Networks
- 2025ICLRShedding Light on Time Series Classification using Interpretability Gated Networks
- 2024NeurIPSAbstracted Shapes as Tokens - A Generalizable and Interpretable Model for Time-series
- 2018ICLRFastGCN: Fast Learning with Graph Convolutional Networks via Importance Sampling