SCALA: Sparsification-based Contrastive Learning for Anomaly Detection on Attributed Networks

TL;DR

SCALA tackles anomaly detection on attributed networks by explicitly handling cases where homophily fails and by mitigating noise from anomalous relationships through graph sparsification. It introduces a dual-view contrastive learning framework (dense-view and spar-view) built on node-graph contrasts, with an attention-based pooling in the sparsified view and a Random Walk with Restart–driven subgraph sampling. The method jointly optimizes dense- and sparse-view losses and infers node anomalies through a fused score that combines contrastive judgments with sparsification signals, evaluated on five real-world networks where anomalies are injected via DOMINANT perturbations. Experimental results demonstrate that SCALA consistently outperforms a broad suite of baselines, highlighting the benefit of sparsification as both a view augmentation and an anomaly signal, thereby improving embedding quality and anomaly scoring in complex networks.

Abstract

Anomaly detection on attributed networks aims to find the nodes whose behaviors are significantly different from other majority nodes. Generally, network data contains information about relationships between entities, and the anomaly is usually embodied in these relationships. Therefore, how to comprehensively model complex interaction patterns in networks is still a major focus. It can be observed that anomalies in networks violate the homophily assumption. However, most existing studies only considered this phenomenon obliquely rather than explicitly. Besides, the node representation of normal entities can be perturbed easily by the noise relationships introduced by anomalous nodes. To address the above issues, we present a novel contrastive learning framework for anomaly detection on attributed networks, \textbf{SCALA}, aiming to improve the embedding quality of the network and provide a new measurement of qualifying the anomaly score for each node by introducing sparsification into the conventional method. Extensive experiments are conducted on five benchmark real-world datasets and the results show that SCALA consistently outperforms all baseline methods significantly.

Figures (5)

• Figure 1: Illustration of the property and the impact of anomalies. (a) The average similarity score of normal and abnormal nodes are calculated by Eq.(1) respectively. We display their percentage distribution by using a histogram. It can be seen that abnormal nodes are more concentrated in areas with low homogeneity. (b) The green and red circle represent the correct and wrong subgraph sampling respectively corresponding to the target node which is presented by the blue dot. We can see that the abnormal node in the subgraph will lead disagreement between the target node and the subgraph embedding.
• Figure 2: The framework of the SCALA.
• Figure 3: ROC curves on five benchmark datasets.
• Figure 4: Performance with different parameters.
• Figure 5: Performance with different $\varepsilon$.