Transfer Learning for Graph Anomaly Detection Using Energy-Based Models

EasyChair Preprint 15464

Abstract

Graph Anomaly Detection (GAD) has applications across social networks, financial systems,
and cyber security. Traditional GAD methods, particularly energy-based models (EBMs),
detect abnormal patterns within graph structures but require extensive training data,
limiting their use on smaller graphs. This paper proposes a novel approach integrating
transfer learning with EBMs to improve anomaly detection performance on graphs with
limited data. The model pre-trains on large source graphs and transfers knowledge to target
graphs with less data, achieving higher accuracy and computational efficiency. We present a
rigorous mathematical foundation and provide detailed experimental results, including
performance metrics, across five tables.

Keyphrases: Algorithms, EBMs, Transfer Learning, graph anomaly detection

@booklet{EasyChair:15464,
  author    = {Lilyana Starlingford and Aarav Thakurani and Mehmmet Amin},
  title     = {Transfer Learning for Graph Anomaly Detection Using Energy-Based Models},
  howpublished = {EasyChair Preprint 15464},
  year      = {EasyChair, 2024}}