Research — Lijing Wang

Research Interests

Building robust, efficient theoretical foundations for large-scale AI models that are reliable when deployed to real-world problems — bridging methodological innovation with high-impact applications.

Methods

Theory-guided machine learning
Deep ensemble algorithms
Efficient learning & model pruning
Spatial-temporal data analysis
Generalization of large language models
Natural language processing

Applications

Recommender systems
User behavior & network analysis
Epidemiological modeling & forecasting
Healthcare analytics with EHR data
Surgical planning & medical imaging
Open innovation & crowdsourcing

Grants

AI for Reliable Epidemic Forecasting

PI · Grace Hopper AI Research Institute · FY25–26

$10,000
Large Language Model-Driven AI Platform for Next-Generation Surgical Planning and Navigation

Co-PI (PI: Zhifeng Kou, Bioinformatics) · Grace Hopper AI Research Institute · FY25–26

$25,000
Provable Efficient Learning with Foundation Models

Co-PI (PI: Shuai Zhang, Data Science) · NJIT Faculty Seed Grant · FY25

$10,000
Towards Improving the Generalization and Robustness of Large Pretrained Language Models

PI (Co-PI: Mengnan Du, Data Science) · NJIT Faculty Seed Grant · FY24

$10,000

Current Research Projects

Generalization, Consistency, and Stability of Large-Scale Models

Generalization · Consistency · Stability

Bridging the gap between training environments and real-world deployment. Large-scale models often exhibit brittle behaviors under adversarial inputs or shifting data streams. We engineer solutions that enforce strict output consistency and improve stability during fine-tuning and lifelong learning, transforming black-box predictors into trustworthy systems for high-stakes domains.

Resource-Efficient Deep Recommendation Systems

Deep Learning · Model Pruning · Feature Selection

Developing efficient deep recommendation systems via model pruning and feature selection. By slashing model complexity and inference latency, this work enables real-time on-device personalization that enhances user experience while safeguarding privacy through reduced data transfer.

GNNs and NLP for Social Science Applications

GNN · NLP · Network Analysis · Behavior Analysis

Applying GNNs and NLP to analyze user behavior and network dynamics in crowdsourcing platforms. By jointly modeling textual data and network interactions, we uncover patterns in collaboration and knowledge-sharing — informing evidence-based design of online communities that foster equitable participation and effective collective problem-solving.

AI in Healthcare: Epidemic Forecasting, Medical Imaging and Surgical Navigation

DL · GNN · NLP · LLM · AI AGENT

In collaboration with bioengineering faculty at NJIT and experts in medical imaging and neural sciences, we are building an AI-based platform for automated surgical planning and navigation, with a focus on brain tumor surgeries. The work leverages LLMs, advanced ML algorithms, and modern imaging techniques to enhance clinician-imaging interaction with precision and efficiency.

Previous Research Projects

Epidemic Forecasting with RNNs and GNNs

RNN · GNN · Dynamic Networks · Mobility Map

Novel frameworks using RNNs and GNNs for spatio-temporal epidemic forecasting. Extensive experiments on seasonal ILI and COVID-19 case datasets demonstrate the value of capturing both temporal dynamics and cross-spatial effects.

Combining Theory and Deep Learning for Epidemic Forecasting

DNN · Theory-Based Causal Models · Synthetic Data

Combining mechanistic causal methods with deep learning to address the black-box nature of pure DL approaches. Models provide correct inference along with explanatory power for underlying epidemic phenomena.

Epidemic Forecasting with Mobility Data

Human Mobility · GNN · Agent-Based SEIR · Metapopulation

Leveraging a large-scale anonymized mobility map aggregated over hundreds of millions of smartphones — factoring it into metapopulation models for retrospective influenza forecasting in the US and Australia, and into GNNs for COVID-19 case forecasting at US state level.

Epidemic Forecasting with Social Media Data

Twitter · Topic Modeling · Agent-Based SEIR

Combining the strengths of social media mining and computational epidemiology — learning individual health status from online posts via Bayesian inference, extracting disease parameters for population-level models, and feeding model outputs back into the mining process for improved real-time prediction.

Health Disparity Analysis via Agent-Based SEIR Simulation

Health Disparity · Vaccination Strategy · Net Return

A computational framework for studying health disparities among cohorts based on individual-level features (age, gender, income). Applied to find disparities in influenza epidemics and to evaluate vaccination prioritization strategies — assisting public health policymakers in efficient allocation of pharmaceutical resources.

Identity Reconciliation via Graph Matching

Social Networks · Percolation-Based Graph Matching

Extending percolation-based graph matching (PGM) — previously limited to undirected networks — to directed online social networks. PDGM incorporates directional relationship similarity and celebrity penalty, outperforming undirected baselines for cross-network identity reconciliation.