FSGAD: Application of Graph Anomaly Detection Based on Multi-View Contrastive Learning in Food Sample Detection
Ruishuang Sun, Chen Chen, Dan Lu, Enguang Zuo, Wei He, Farong Chen · 2025
Food safety is a major global concern, especially nowadays when food samples are increasingly complex. However, traditional detection methods only rely on a single type of data, which makes it difficult to model complex feature associations among multiple samples. This paper proposes a multi-view contrastive learning method for food sample detection (FSGAD). A novel contrastive instance pair is used to capture the relationship between the target node and its neighbors and non-neighbors. Starting from the node and line graph, utilize the complementary information between them to capture anomalies. In order to effectively handle different types of graph data, a contrastive learning method combining batch graphs (integrated line graphs) and single graphs (processed node graphs) is designed to learn information embedding from DGL graph attributes and the local structure of the graphs, respectively. Finally, Binary cross-entropy loss is used to measure the difference between the true and predicted values. This study provides an innovative technical approach to food safety and lays the foundation for further improving the reliability of food detection.