Abstract
Networks are ubiquitous in the real world such as social networks and communication networks, and anomaly detection on networks aims at finding nodes whose structural or attributed patterns deviate significantly from the majority of reference nodes. However, most of the traditional anomaly detection methods neglect the relation structure information among data points and therefore cannot effectively generalize to the graph structure data. In this paper, we propose an end-to-end model of Deep Dual Support Vector Data description based Autoencoder (Dual-SVDAE) for anomaly detection on attributed networks, which considers both the structure and attribute for attributed networks. Specifically, Dual-SVDAE consists of a structure autoencoder and an attribute autoencoder to learn the latent representation of the node in the structure space and attribute space respectively. Then, a dual-hypersphere learning mechanism is imposed on them to learn two hyperspheres of normal nodes from the structure and attribute perspectives respectively. Moreover, to achieve joint learning between the structure and attribute of the network, we fuse the structure embedding and attribute embedding as the final input of the feature decoder to generate the node attribute. Finally, abnormal nodes can be detected by measuring the distance of nodes to the learned center of each hypersphere in the latent structure space and attribute space respectively. Extensive experiments on the real-world attributed networks show that Dual-SVDAE consistently outperforms the state-of-the-arts, which demonstrates the effectiveness of the proposed method.
Introduction
Attributed networks, where nodes are characterized by both their individual attributes and their structural relationships, are common in various domains like social media and communication systems. Anomaly detection in these networks is critical for identifying entities that deviate from the norm. These deviations typically manifest in three forms: structure anomalies, attribute anomalies, and combined anomalies.
Traditional methods often struggle with these complex patterns because they either ignore the graph structure or rely solely on reconstruction errors, which can be sensitive to noise and fail to capture the nuanced interplay between a node's content and its context.
Methodology: The Dual-SVDAE Framework
The proposed Deep Dual Support Vector Data Description based Autoencoder (Dual-SVDAE) addresses these challenges through a two-pronged architecture designed to learn distinct yet fused representations of structure and attributes.
Dual Autoencoder Architecture
The model comprises two main components:
- Structure Autoencoder: Utilizes a Graph Convolutional Network (GCN) to process the adjacency matrix, capturing the relational context of each node.
- Attribute Autoencoder: Employs a Multi-Layer Perceptron (MLP) to extract features from the raw node attributes.
Dual-Hypersphere Learning
A key innovation is the integration of Support Vector Data Description (SVDD) principles. Instead of relying only on reconstruction, the model maps "normal" nodes into two compact hyperspheres—one for the structure space and one for the attribute space. This explicitly defines a boundary for normalcy.
Feature Fusion and Decoding
To ensure the model captures the joint relationship between structure and attributes, the latent embeddings from both spaces are fused. This fused representation is then used by a decoder to reconstruct the node attributes, forcing the model to learn a comprehensive representation.
Anomaly Scoring
Anomalies are detected by calculating a weighted distance from the centers of the two learned hyperspheres. Nodes falling outside these boundaries are assigned higher anomaly scores.
Experiments and Results
The Dual-SVDAE model was evaluated on three benchmark citation networks: Cora, Citeseer, and Pubmed. The experiments followed a one-class classification setup where one class was designated as "normal" and others as anomalies.
The results demonstrate that Dual-SVDAE consistently outperforms state-of-the-art baselines:
- Cora: Achieved an AUC of 95.09%, significantly higher than OC-SAGE (90.99%).
- Pubmed: Reached an AUC of 91.32%, showing a substantial improvement over traditional methods that often failed to exceed 75%.
Ablation studies confirmed that both the dual-hypersphere mechanism and the feature fusion module are essential for high performance. Visualizations using t-SNE further illustrated the model's ability to tightly cluster normal nodes while isolating anomalies at the periphery.
Conclusion
Dual-SVDAE introduces a robust approach to network anomaly detection by combining deep autoencoders with dual-hypersphere learning. By explicitly modeling boundaries of normalcy in both structure and attribute spaces, it provides a more effective and interpretable detection mechanism than traditional reconstruction-based models. Future research will focus on adapting this framework for dynamic networks and automating threshold selection for broader industrial applicability.
Summary updated