FFRAUD‐ER: Development of a computational model for identifying Food Fraud incidents as drivers for Food Safety Emerging Risks

Abstract The primary objective of this project was the development of a computational model aimed at identifying food fraud incidents as drivers for food safety emerging risks, in alignment with EFSA's activities on environmental scanning. These activities include the identification and analysis of emerging risks, the enhancement of risk identification methodologies, and the communication of identified risks. At first stage, the focus was on the identification and prioritisation of data sources of historical food fraud cases. This phase involved rigorous criteria for data selection to ensure accuracy and relevance. Subsequently, a labelled dataset was developed through systematic data categorisation and annotation, essential for training the computational model to detect patterns of potential safety risks in food fraud incidents. The core of the project involved the creation of the Computational Model (CM) using natural language processing (NLP) and deep learning algorithms. Six state‐of‐the‐art NLP models (BERT, XLNet, GPT‐3, ELECTRA, T5, and ELMo) were identified, trained and tested on the labelled dataset, with XLNet achieving an F1‐Score of 95% and a ROC/AUC Score of 98%. An iterative refinement phase followed, involving feedback from the FFRAUD‐ER Network and stakeholders to enhance the model's reliability. Finally, the model's application in real‐world scenarios was explored, demonstrating its utility in aiding EFSA and other Emerging Risk Identification networks to detect emerging food safety risks. The results indicate that the computational model can successfully determine whether a food fraud instance poses a food safety risk. Statistical and analytical techniques applied alongside the CM on newly acquired data of food fraud incidents have shown potential, although their efficiency and effectiveness will primarily depend on the quality and quantity of future data.