在屬性圖中進行子圖層次異常檢測之結構感知雙對比圖神經網絡

Abstract

在網絡中識別異常子圖對於各種應用至關重要，包括疾病爆發檢測、金融欺詐檢測和社交網絡活動監控。然而，在屬性和結構上存在多樣性異常的情況下識別這些異常子圖非常具有挑戰性。雖然圖神經網絡近來提供了一種端到端的方法來學習異常度量函數，但由於子圖表示學習和非監督異常量化方面存在重大挑戰，仍需要探索子圖層級異常的檢測。此外，現有方法通常采用自下而上的方法，訓練節點（或邊）檢測模型，將異常實體聯繫起來形成子圖。然而，這些方法可能忽略全局信息，其中異常子圖中的節點可能是普通的，但與其他子圖相比被視為異常。為解決這一問題，本文介紹了一種新的圖神經網絡框架，稱為結構感知雙對比圖神經網絡（SDCGNN），以自上而下的方式提取異常子圖。SDCGNN包括兩個主要組件：分層分組模塊（HGM），利用結構信息和屬性與結構之間的協同資訊將輸入圖劃分為子圖；基於對比的異常子圖檢測模塊（CASDM），區分異常子圖和正常子圖。對真實和合成數據集的廣泛實驗結果證實了SDCGNN在具有結構和屬性異常的情況下相對於十個基線的優越性能。

Identifying anomalous subgraphs within networks is vital for various applications, including disease outbreak detection, financial fraud detection, and social network activity monitoring. However, the challenge lies in recognizing these anomalous subgraphs, given the diverse anomalies in attributes and structure. While graph neural networks have recently offered an end-to-end approach to learning anomaly measure functions, the detection of subgraph-level anomalies still needs to be explored due to significant challenges in subgraph representation learning and unsupervised anomaly quantification. Moreover, existing methods often adopt a bottom-up approach, training node (or edge) detection models and linking abnormal entities to form a subgraph. However, these approaches may neglect global information, where nodes in an anomalous subgraph might be average yet considered anomalies compared to other subgraphs. To address this, the paper introduces a novel Graph Neural Network framework called Structure-aware Dual Contrastive Graph Neural Network (SDCGNN) to extract anomalous subgraphs in a top-down manner. SDCGNN incorporates two key components: the Hierarchical Grouping Module (HGM), dividing the input graph into subgraphs using both structural information and collaborative data between attributes and structure, and the Contrastive-based Anomalous Subgraph Detection Module (CASDM), distinguishing anomalous subgraphs from normal ones. Extensive experimental results on real and synthetic datasets substantiate its superior performance over ten baselines with structural and attribute abnormalities.