In recent years, edible oil safety problems have occurred frequently. In order to reduce the threat of such incidents, it is of great significance to research edible oil safety risk assessment models. Considering that high-dimensional, non-linear and discrete data containing noise are currently obtained from the detection of edible oils, and the existing risk assessment models have several problems such as poor noise suppression, inaccurate evaluation, and strong subjectivity in model parameter adjustment, a risk assessment model for pollutants in edible oils was proposed in this paper. First, risk indicators were selected and data were preprocessed and input into a filtering module based on the wavelet threshold method for filtering. Second, grey relational analysis (GRA) was used to calculate the weight of each risk index and develop a multi-index comprehensive risk label. Extreme learning machine (ELM) was adopted to predict the comprehensive risk value. Third, the practical Bayesian optimization (PBO) algorithm was used to optimize the parameters of filtering module and ELM network. Finally, the fuzzy comprehensive analysis was applied to classify the risk grade of the predicted comprehensive risk value. The application of the proposed model to 150 groups of edible oil data was described in detail. The coefficient of determination (R²) and root mean square error (RMSE) of this model were 0.0563 and 0.9461, respectively, indicating its superiority and effectiveness. This study provides reasonable evidence for relevant departments to formulate risk control and sample inspection strategies and optimize the supply chain of edible oils.

In order to solve the problems of strong subjectivity and low targeting in sampling decision-making that exist in food safety surveillance sampling, this study proposed a correlation analysis method based on an improved Frequent Pattern-growth (FP-growth) algorithm for food risk factors. First, the entropy weight method was used to assign weights to the risk indicators of food categories so as to calculate the risk indices of different food categories. Second, the risk index was used as a feature for risk clustering based on MiniBatchKmeans to obtain the risk level of food products. Finally, an improved FP-growth algorithm with constraints was used for association rule mining of food risk factors to excavate the association relationship between the risk level of food products and the information of food types, time, and geographic attributes, and the mined results were analyzed by correlation analysis so as to provide guidance for precise targeting to guide the decision making of sampling inspection. This study was based on food sampling data from certain regions of China in 2019, which were assigned with indicators to calculate the risk index. Afterwards, the risk was clustered into low (L), medium (M), and high risk (H). Finally, the data was imported into the improved FP-growth algorithm to obtain the association rules of food risk factors. For 17214 pieces of sampling data, the improved FP-growth algorithm had a shorter running time when compared with the Apriori algorithm. Compared with the traditional one, the improved FP-growth algorithm removed invalid rules and improved the analysis efficiency of the association rules of food risk factors. Thus, it provides an accurate and efficient decision-making basis for the sampling work of food regulatory authorities.