Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Anomaly-Aware Pre-Training and Fine-Tuning for Graph Anomaly Detection

Yunhui Liu, Jiashun Cheng, Yiqing Lin, Qizhuo Xie, Jia Li, Fugee Tsung, Hongzhi Yin, Tao Zheng, Jianhua Zhao, Tieke He

TL;DR

This work tackles graph anomaly detection under severe label scarcity and local homophily disparity by introducing Anomaly-Aware Pre-Training and Fine-Tuning (APF). APF combines Rayleigh Quotient-guided subgraph sampling with dual spectral filters to learn anomaly-relevant representations during pre-training, and a granularity-adaptive fusion mechanism with anomaly-aware regularization during fine-tuning. Theoretical analysis in an anomalous stochastic block model framework supports linear separability under mild conditions, while extensive experiments on 10 GADBench datasets demonstrate strong, robust performance across diverse domains. The approach offers a scalable, interpretable pathway to effective GAD under realistic supervision constraints, with clear mechanisms to adapt to node- and dimension-level heterogeneity.

Abstract

Graph anomaly detection (GAD) has garnered increasing attention in recent years, yet remains challenging due to two key factors: (1) label scarcity stemming from the high cost of annotations and (2) homophily disparity at node and class levels. In this paper, we introduce Anomaly-Aware Pre-Training and Fine-Tuning (APF), a targeted and effective framework to mitigate the above challenges in GAD. In the pre-training stage, APF incorporates node-specific subgraphs selected via the Rayleigh Quotient, a label-free anomaly metric, into the learning objective to enhance anomaly awareness. It further introduces two learnable spectral polynomial filters to jointly learn dual representations that capture both general semantics and subtle anomaly cues. During fine-tuning, a gated fusion mechanism adaptively integrates pre-trained representations across nodes and dimensions, while an anomaly-aware regularization loss encourages abnormal nodes to preserve more anomaly-relevant information. Furthermore, we theoretically show that APF tends to achieve linear separability under mild conditions. Comprehensive experiments on 10 benchmark datasets validate the superior performance of APF in comparison to state-of-the-art baselines.

Towards Anomaly-Aware Pre-Training and Fine-Tuning for Graph Anomaly Detection

TL;DR

This work tackles graph anomaly detection under severe label scarcity and local homophily disparity by introducing Anomaly-Aware Pre-Training and Fine-Tuning (APF). APF combines Rayleigh Quotient-guided subgraph sampling with dual spectral filters to learn anomaly-relevant representations during pre-training, and a granularity-adaptive fusion mechanism with anomaly-aware regularization during fine-tuning. Theoretical analysis in an anomalous stochastic block model framework supports linear separability under mild conditions, while extensive experiments on 10 GADBench datasets demonstrate strong, robust performance across diverse domains. The approach offers a scalable, interpretable pathway to effective GAD under realistic supervision constraints, with clear mechanisms to adapt to node- and dimension-level heterogeneity.

Abstract

Graph anomaly detection (GAD) has garnered increasing attention in recent years, yet remains challenging due to two key factors: (1) label scarcity stemming from the high cost of annotations and (2) homophily disparity at node and class levels. In this paper, we introduce Anomaly-Aware Pre-Training and Fine-Tuning (APF), a targeted and effective framework to mitigate the above challenges in GAD. In the pre-training stage, APF incorporates node-specific subgraphs selected via the Rayleigh Quotient, a label-free anomaly metric, into the learning objective to enhance anomaly awareness. It further introduces two learnable spectral polynomial filters to jointly learn dual representations that capture both general semantics and subtle anomaly cues. During fine-tuning, a gated fusion mechanism adaptively integrates pre-trained representations across nodes and dimensions, while an anomaly-aware regularization loss encourages abnormal nodes to preserve more anomaly-relevant information. Furthermore, we theoretically show that APF tends to achieve linear separability under mild conditions. Comprehensive experiments on 10 benchmark datasets validate the superior performance of APF in comparison to state-of-the-art baselines.

Paper Structure

This paper contains 57 sections, 2 theorems, 23 equations, 13 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Preliminary
  3. Graph Anomaly Detection (GAD).
  4. Local Homophily.
  5. Graph Spectral Filtering.
  6. Methodology
  7. Anomaly-Aware Pre-Training
  8. Label-free Anomaly Indicator.
  9. Dual-filter Encoding.
  10. Optimization.
  11. Granularity-Adaptive Fine-Tuning
  12. Node- and Dimension-wise Fusion.
  13. Anomaly-aware Regularization Loss.
  14. Optimization.
  15. Theoretical Insights
  16. ...and 42 more sections

Key Result

Lemma 1

The Rayleigh Quotient $RQ(x, L)$, which represents the accumulated spectral energy of a graph signal, increases monotonically with the anomaly degree. BWGNN

Figures (13)

  • Figure 1: (a), (b): Distribution of local homophily for Weibo and T-Finance. (c), (d): Performance across local homophily quartiles (Q1 = top 25%, Q4 = bottom 25%) on Weibo and T-Finance.
  • Figure 2: Overview of our proposed APF.
  • Figure 3: Performance variations across local homophily quartiles (Q1 = top 25%, Q4 = bottom 25%).
  • Figure 4: Visualization of the learned coefficients for the top 6 dimensions. The nodes $a_1, a_2, a_3$ are 3 randomly selected abnormal nodes, while $n_1, n_2, n_3$ are 3 randomly selected normal nodes.
  • Figure 5: How the AUPRC score varies with different values of $p_a$ and $p_n$.
  • ...and 8 more figures

Theorems & Definitions (4)

  • Lemma 1
  • Definition 1: $ASBM(n_a, n_n, \boldsymbol{\mu}, \boldsymbol{\nu}, (p_1, q_1), (p_2, q_2), P_a, P_n)$
  • Theorem 1
  • proof