Evaluation of methods for the detection of hazardous substances in food based on machine learning

Abstract

With new detection methods for hazards in food being created, it is essential to build an evaluation model of these methods in order to improve the level and efficiency of food safety related detection. Back propagation neural network (BPNN), classification and regression tree (CART), support vector regression (SVR) and random forest regression (RFR) are the supervised machine learning methods used to build the model. In this study, we used these four approaches to evaluate hazard detection methods. First, the factors that affect the test (considering the accuracy, efficiency and cost of the methods) were considered, and 8 indices were selected to establish an evaluation index system for chemical hazards and biological hazards respectively. Then, 727 methods related to the detection of chemical hazards and 91 methods related to the detection of biological hazards in food were analyzed, and experts were invited to score each method as machine learning training data. Then, four machine learning models were used for training and verification. It turns out that for the detection methods of chemical hazards, the BPNN, CART and SVR models require a shorter period of time to adjust parameters, perform better on the test set with great fitting ability and lower mean square error, and the goodness of fit is 0.969, 0.962 and 0.988, respectively. While for the detection methods of biological hazards, BPNN and SVR models fit better with goodness of fit values of 0.988 and 0.999, respectively, and give more accurate prediction. It can be concluded that BPNN and SVR models perform well for both chemical and biological hazards detection methods, which can be used as a reference for artificial selection before testing.