Yuxiang Wang

Yuxiang Wang

MS @ Fudan University. Research interests: LLMs (Agents), Diffusion Language Models, AI4S, Multimodal LLMs, GNNs.

STM: Spatio-Temporal Distance and Frame-Based Dynamic Graph Fraud Detection

Published in CCF-B venue (Springer), 2024

Motivation

Conventional GNN-based fraud detectors struggle to fully exploit the rich spatio-temporal structure present in dynamic graphs, leaving long-range temporal signals and inter-frame discrepancies underused.

Method — STM

STM decomposes the dynamic graph into frames and applies multiple complementary modules:

  • Graph condensation — compresses the per-frame substructure into compact, learnable summaries.
  • Distance-aware intra-frame aggregation — aggregates neighborhood signals weighted by spatio-temporal distance.
  • Difference-aware inter-frame aggregation — captures behavioral drift across frames to surface emerging fraud patterns.

Together these modules let STM exploit both short-term and long-term temporal information alongside spatial structure, improving both detection accuracy and efficiency.

Outcomes

  • SOTA on multiple dynamic-graph fraud-detection benchmarks
  • Granted invention patent
  • Funded as a National-level College Student Innovation Project (lead PI)

Recommended citation: Yuxiang Wang, et al. "Spatio-Temporal Distance and Frame-Based Dynamic Graph Fraud Detection." Springer, 2024. https://doi.org/10.1007/978-981-95-3906-2_3
Download Paper