A Pre-Training and Adaptive Fine-Tuning Framework for Graph Anomaly Detection

2025-04-19

AI Summary

Graph anomaly detection (GAD) faces three key challenges: severe scarcity of anomalous samples, high annotation cost, and the coexistence of local heterogeneity in anomalous nodes and global homogeneity in normal nodes—a “mixed homogeneity/heterogeneity spectrum” that existing graph pretraining methods fail to capture. To address this, we propose a spectral-aware pretraining–adaptive fine-tuning framework. First, we theoretically characterize how mixed homogeneity and heterogeneity jointly shape graph filter responses in the spectral domain. Second, we design a node-level adaptive dual-pass filtering mechanism that abandons the restrictive single low-pass assumption. Third, we integrate low-pass and high-pass graph filtering, a gated fusion network, and spectral-analysis-driven representation learning. Evaluated on ten benchmark datasets, our method consistently outperforms state-of-the-art approaches, achieving average AUC gains of 3.2–7.8 percentage points—demonstrating both the effectiveness and generalizability of spectral adaptive modeling for GAD.

Technology Category

• Machine Learning: Graph-based Machine Learning
• Data Mining & Knowledge Management: Graph Mining, Social Network Analysis & Community
• Reasoning under Uncertainty: Graphical Models

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Graph anomaly detection (GAD) has garnered increasing attention in recent years, yet it remains challenging due to the scarcity of abnormal nodes and the high cost of label annotations. Graph pre-training, the two-stage learning paradigm, has emerged as an effective approach for label-efficient learning, largely benefiting from expressive neighborhood aggregation under the assumption of strong homophily. However, in GAD, anomalies typically exhibit high local heterophily, while normal nodes retain strong homophily, resulting in a complex homophily-heterophily mixture. To understand the impact of this mixed pattern on graph pre-training, we analyze it through the lens of spectral filtering and reveal that relying solely on a global low-pass filter is insufficient for GAD. We further provide a theoretical justification for the necessity of selectively applying appropriate filters to individual nodes. Building upon this insight, we propose PAF, a Pre-Training and Adaptive Fine-tuning framework specifically designed for GAD. In particular, we introduce joint training with low- and high-pass filters in the pre-training phase to capture the full spectrum of frequency information in node features. During fine-tuning, we devise a gated fusion network that adaptively combines node representations generated by both filters. Extensive experiments across ten benchmark datasets consistently demonstrate the effectiveness of PAF.

Problem

• Addressing graph anomaly detection with scarce abnormal nodes
• Overcoming limitations of global low-pass filters in GAD
• Adaptively combining filters for complex homophily-heterophily mixtures

Innovation

• Joint training with low-high pass filters
• Adaptive gated fusion network
• Pre-training and fine-tuning framework