A comparison of neural and non-neural machine learning models for food safety risk prediction with European Union RASFF data

European Union launched the RASFF portal in 1977 to ensure cross-border monitoring and a quick reaction when public health risks are detected in the food chain. There are not enough resources available to guarantee a comprehensive inspection policy, but RASFF data has enormous potential as a preventive tool. However, there are few studies of food and feed risk issues prediction and none with RASFF data. Although deep learning models are good prediction systems, it must be confirmed whether in this field they behave better than other machine learning techniques. The importance of categorical variables encoding as input for numerical models should be specially studied. Results in this paper show that deep learning with entity embedding is the best combination, with accuracies of 86.81%, 82.31%, and 88.94% in each of the three stages of the simplified RASFF process in which the tests were carried out. However, the random forest models with one hot encoding offer only slightly worse results, so it seems that in the quality of the results the coding has more weight than the prediction technique. Our work also demonstrates that the use of probabilistic predictions (an advantage of neural models) can also be used to optimize the number of inspections that can be carried out. •It has been created a training set formed by historical RASFF data.•The best predictive model uses categorical embeddings plus deep learning techniques.•A comparison between neural and non-neural machine learning models has been done.•Three different deep learning models are developed with accuracies from 72 to 81%.