Research

Research interests include:

• General: natural language processing, machine learning, deep learning, sparse optimization, large-scale optimization, mathematical finance, mathematical modeling, quantum computing, numerical linear algebra
• Methods and topics: support vector machines, penalty decomposition, L1-minimization, greedy algorithms, semidefinite programming, ADMM, inexact Newton methods, matrix decomposition, linear programming, and related areas

Current research projects

1. Soft hate speech detection via hierarchical event-aware modeling (HEART-Soft)

   This project introduces HEART-Soft, a framework designed to detect subtle and implicit forms of harmful language by modeling discourse structure, contextual semantics, and temporal evolution. Instead of analyzing posts in isolation, the approach groups them into events and processes them through overlapping temporal windows. It integrates contextual retrieval, semantic inconsistency signals, and hierarchical attention mechanisms to identify soft hate expressions that may not be explicitly stated. Additionally, the model provides interpretability through risk factors, contextual attribution, and uncertainty estimation, making it suitable for real-world deployment in sensitive settings. This work contributes to responsible AI and nuanced language understanding.

2. Exact sparse optimization via penalty decomposition and safeguarded MM

   This project develops a general optimization framework for ℓ₀-constrained problems under polyhedral and structured constraints. The methodology combines penalty decomposition (PD) with majorization–minimization (MM) techniques, enabling the separation of discrete sparsity enforcement from continuous optimization. By incorporating safeguarding strategies and flexible surrogate models (e.g., diagonal and quasi-Newton approximations), the approach remains stable even in nonconvex and ill-conditioned settings. Theoretical results target convergence to Lu–Zhang stationary points, while practical implementations emphasize efficiency and robustness. This project serves as a core theoretical foundation for multiple applications.

3. Event-centric multimodal misinformation detection (E-CaTCH)

   This project develops E-CaTCH, a framework for misinformation detection that integrates textual, visual, and temporal signals within an event-based structure. By clustering posts into events and temporal sequences, the model captures how misinformation evolves and propagates over time. It employs intra-modal and cross-modal attention mechanisms, along with temporal modeling via sequence models, to produce robust predictions. The framework also incorporates class imbalance handling and interpretability mechanisms, making it both effective and transparent. This work contributes to multimodal learning and trustworthy AI systems.

4. AI-assisted mathematical discovery system

   This project aims to build an AI system that mirrors the process of mathematical discovery, including intuition formation, experimental validation, and conjecture generation. The system autonomously generates hypotheses, runs computational experiments, and iteratively refines conjectures based on observed patterns. By structuring discovery as a sequence of optimization and validation steps, the framework significantly reduces the time between initial insight and rigorous testing. The long-term vision is to position AI as a collaborator in mathematical research, capable of exploring large conceptual spaces and assisting in theory development.