Mingjian Lu, Hieu Vu, Vu K Le, Jing Ma, Yinghui Wu

CATGX: Causal-Aware Temporal Graph Explanation via Scalable Motif Sampling and Adjustment.

The 32nd SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2026

Mingjian Lu, Haolai Che, Yangxin Fan, Qu Liu, Fei Shao, Tingjian Ge, Xusheng Xiao, Yinghui Wu.

Training-free Counterfactual Explanation for Temporal Graph Model Inference.

ICLR 2026 

Yangxin Fan, Haolai Che, Mingjian Lu, Yinghui Wu.

Graph Compression for Interpretable Graph Neural Network Inference at Scale.

PVLDB 2025

Hernandez, K. J., Ciardi, T. G., Yamamoto, R., Mingjian Lu, et al.

L-PBF High-Throughput Data Pipeline Approach for Multi-Modal Integration.

Integrating Materials and Manufacturing Innovation, 2024

Mingjian Lu, Chen, Q.Q., Chen, Y.J., Sun, W.J.

Micromanagement in StarCraft Game AI: A Case Study.

Procedia Computer Science (IIKI), 2019

Physics-Grounded Multimodal AI — AAAI 2027 Under Review

• Built an end-to-end multimodal AI pipeline combining image segmentation, vision encoders, attention-based fusion, and LLM-generated explanations for scientific imaging.

• Integrated domain constraints into multimodal representations to generate grounded, domain-consistent explanations rather than relying solely on black-box model predictions.

• Benchmarked modern vision foundation models and segmentation architectures, including JEPA, DINOv2, SAM, SegFormer, and DeepLabv3+, across multiple imaging domains.

• Designed a physics-fidelity evaluation metric to quantify alignment between model explanations and domain knowledge while preserving predictive faithfulness. 

Causal & Scalable Temporal Graph Explanation — CATGX | KDD 2026

• Built a causal explanation system for temporal graph neural networks, separating true structural effects from confounding signals such as recency, duration, and user/entity activity.

• Designed a motif-level causal attribution pipeline with backdoor adjustment to produce more reliable explanations for large-scale dynamic graph predictions.

• Engineered an HNSW-based approximate-nearest-neighbor sampling system with stratified importance weighting, reducing explanation runtime to 48 seconds on million-edge graphs.

• Delivered up to 83× speedup over existing baselines while maintaining 0.85–0.90 motif coverage across production-scale temporal graph benchmarks. 

Training-Free Temporal Graph Explanation — TemGX | ICLR 2026

• Built TemGX, a training-free counterfactual explanation system for temporal graph models, supporting both node classification and link prediction without retraining production models.

• Designed temporal influence and graph-distance scoring algorithms to identify decision-critical events and substructures behind model predictions.

• Evaluated the system on traffic, financial transaction, and dynamic-network datasets, achieving strong fidelity and sparsity on METR-LA, PEMS-BAY, and Bitcoin data.

• Integrated model explanations with multi-layer knowledge graphs to support investigation of evolving patterns such as transaction chains, attack stages, and material failures. 

Multimodal Graph AI for Failure Prediction

• Built a multimodal vision–graph learning pipeline that converts raw SEM images into structured defect graphs for downstream failure prediction.

• Developed an automated SegFormer-B2 → graph construction → GraphSAGE pipeline, representing detected defects as nodes and spatial/material relationships as graph edges.

• Designed dual GraphSAGE models to learn complementary representations from fatigue and overload regions for microstructure-aware prediction.

• Generated visual and graph counterfactual explanations and grounded LLM outputs with knowledge graphs, enabling domain experts to trace predictions back to image-level evidence