Search

Food safety is a global issue; risk prevention and control of food physical and chemical testing (PCT) data are critical. However, existing studies have not considered the risk status of qualified products near the detection limit. Therefore, this paper presents a data-driven paradigm to detect the combined risk of multiple indicators across products in real-world food safety datasets for the first time. The following challenges exist to achieve this goal. (1) In real-world scenarios, neighboring samples are somewhat correlated; however, capturing and quantifying the actual testing sample correlation network remains an open problem in food computing. (2) The low-frequency nature of the unqualified samples leads to label imbalance in the data. (3) Labeling dependency limits the application scenarios. This article proposes a novel self-supervised framework (SUCOLA) to address the abovementioned challenges. That jointly learns the optimal graph topology and also inferences the connections between every product node and its well-designed subgraph, helping to resist label imbalance. We applied it to a real-world sterilized PCT dataset from a province in China. Extensive experiments show that SUCOLA performance is outstandingly better than several baselines of the current graph self-supervision field. Meanwhile, we conducted generalizability experiments on five well-known benchmark datasets. The proposed SUCOLA surpasses the current state-of-the-art by a considerable margin (0.39%–8.7% AUC) on public datasets. Thus, our study provides a new idea for early warning research on PCT data in food computing and fills the gap in the study of structural learning under category imbalance. [ABSTRACT FROM AUTHOR]