Un système explicable de recommandation alimentaire multi-LLM basé sur l'apprentissage profond

Abstract

Food recommender systems (FRS) increasingly support dietary decision-making, yet their explanations often remain rigid, poorly contextualized, and disconnected from user needs. Current XAI techniques such as SHAP or LIME provide numerical justifications that lack semantic depth, while single-LLM pipelines restrict adaptability and may amplify model-specific biases. To address these limitations, we introduce a multi-LLM explainability overlay that generates contextual, contrastive, counterfactual, and simulation-based explanations over existing deep learning models. The system integrates Food and User Ontologies, partly derived from the Food Explanation Ontology (FEO) and extended in this work to better represent African and multicultural dietary contexts. The architecture decouples prediction from explanation generation and leverages four LLMs : GPT-4, Gemini, LLaMA-3, and Mistral with a conflict-resolution mechanism and safeguards to ensure fidelity to the underlying model and reduce hallucinations. Experiments on AllRecipesExtended and Food.com datasets show that the proposed system delivers more adaptive and personalized explanations with competitive clarity and significantly lower latency compared to a single-LLM baseline. User feedback further highlights the value of ontology-guided reasoning and culturally contextualized narratives. This work opens a pathway toward more transparent, robust, and human-centered explainability in personalized nutrition.