Attribute-incomplete graph anomaly detection network

doi:10.1016/j.patcog.2026.114173

Attribute-incomplete graph anomaly detection network

2026-06-01

1

PRE

AI

DOI：10.1016/j.patcog.2026.114173

Original

Original request for help

Save

Abstract

En 中文

Anomaly detection with attribute-incomplete graphs is a critical yet challenging issue, prevalent across various domains, yet existing methods have not solved it. When node attributes are incomplete, detection becomes disabled. One possible solution is node attribute imputation followed by detection, but the imputation step typically focuses on restoring general adjacency or distribution information, neglecting detection-friendly features. In addition, the detection process is not optimized for imputed node attributes, making it difficult to extract subtle anomaly patterns effectively. To address this challenge, we propose an Attribute-IncoMplete Graph Anomaly Detection Network (AMGA), which alternately performs node attribute inference and anomaly detection, where the inferred attributes enhance detection accuracy; in turn, the detection results refine the quality of imputation. Specifically, we design the Rayleigh-Guided Attribute Imputation (RGAI) mechanism that extracts reliable Rayleigh subgraphs to drive spectral filtering for node attribute inference. By combining local topology and confidence priors, RGAI ensures smooth imputations in the normal areas while highlighting anomalies. Additionally, we develop a specialized Inference-Aware Anomaly Detection (IAD) module, which leverages the dual-channel representations of community partitions and prior imputation signals to amplify the structural and representation space deviations of anomalous nodes, significantly enhancing the ability to distinguish different patterns. Extensive experiments demonstrate that the AMGA consistently achieves superior detection performance, even under severe node attribute incompleteness, outperforming existing advanced detection methods for graphs with complete attributes. Our source code is available at: https: //github.com/H00001/AMGA-code.

Keywords:

Graph anomaly detection

GRaph Neural Network

Unsupervised learning

Attribute-incomplete graph anomaly detection network

Abstract

Journal

Organization

Cited Papers