We study how large language models reason and how to make that reasoning faithful and useful in high-stakes settings. This includes knowledge-enhanced agentic causal discovery for interpretable and interactable predictive healthcare [EMNLP 25], intermediate-variable reasoning (EMNLP 26), temporal reasoning over clinical narratives (SDM 26), LLMs for socio-political event interpolation and extrapolation (ECML-PKDD 26), and probing the conceptual hierarchies represented inside LLMs (ACL Findings 26).

We design new graph neural networks for dynamic and heterogeneous graph data [KDD 19, KDD 20]; We also work on continual graph learning when tasks/data distributions change over time; We focus on novel solutions for knowledge graph reasoning, graph fusion, and temporal graph predictions. We study causality enhanced machine learning to improve interpretability [ICDM 22, KDD 22], continual graph learning [NeurIPS 24], graph out-of-distribution generalization (AISTATS 25), and graph unlearning [KDD 23].

We design new transfer learning and multitask learning methods for domain adaptation and bias-mitigation [ICWSM 20]. We also develop multitask learning for imbalanced data and spatiotemporal prediction problems [SDM 18]. We investigate federated learning frameworks for asynchronous settings [BigData 20] and heterogeneous data. We also apply these ideas to science and economics, including cross-reservoir inflow prediction [ICDM-DMESS 25, AAAI-AI4ES 26] and inflation forecasting with dynamic feature spaces (ICDM 25).

Working with clinical collaborators at Yale School of Medicine, UMass Chan Medical School, and Hackensack Meridian Health, we develop accurate and transparent models for clinical prediction and diagnostic reasoning, including concept-grounded domain adaptation [ICML 26], clinical domain generalization (CIKM 26), lab-informed pretraining for diagnosis [J-BHI 26], and contrastive ICD coding [NeurIPS 23]. Earlier, we utilized health data to develop new machine learning algorithms for personalized care [AAAI 22] and epidemic forecasting [CIKM 20]. We design new domain knowledge guided deep learning models [IJCAI 21] to discover patient-disease relations, hidden disease patterns, and disease topographies. We focus on tasks such as future risk assessment, ICD coding, medical representation learning, and information retrieval.

We design deep neural networks for societal event predictions including crime, political events, and pandemics. We study new methods for integrating multimodal data [CIKM 21] and causal inference [ICDM 22, KDD 22] in human event analysis.

We are interested in developing efficient and effective detection approaches for socially responsible AI which includes knowledge-based fake news detection [PAKDD 21], fairness in finance [TheWebConf 20], medicine [eBioMedicine], and social networks [ICWSM 22a], and toxic/hate speech detection[ICWSM 20, ICWSM 22b].