Symbolic knowledge extraction for explainable nutritional recommenders.

Area of Science:

• Artificial Intelligence
• Computer Science
• Nutrition Science

Background:

• Nutritional recommendation systems (RS) aim to help users achieve body shape goals.
• Designing effective RS requires balancing expert prescriptions, user preferences, and recommendation explainability.
• Current systems often struggle to integrate these potentially conflicting requirements.

Purpose of the Study:

• To propose a novel approach for engineering nutritional RS that harmonizes user preferences, expert advice, and explainability.
• To develop a system that generates personalized and nutritionally sound dietary recommendations.
• To ensure the transparency of the recommendation generation process.

Main Methods:

• Utilized neural networks (NN) to predict user preferences from recipe data.
• Employed Classification and Regression Trees (CART) to extract symbolic rules from NN predictions.
• Integrated extracted rules with expert prescriptions in Prolog format for a symbolic knowledge base.
• Used logic solvers to query the knowledge base for generating explainable recommendations.

Main Results:

• Neural networks achieved approximately 86% test-set accuracy in predicting user preferences.
• Extracted symbolic rules demonstrated about 80% fidelity with the trained neural networks.
• The developed recommendation system achieved a test-set precision of approximately 74%.
• The symbolic approach enabled clear explanations of how recommendations are generated.

Conclusions:

• The proposed approach effectively integrates user preferences and expert nutritional guidance into intelligent agents.
• The system generates recommendations that are both user-acceptable and nutritionally adequate.
• The method enhances the explainability of AI-driven nutritional recommendations, fostering user trust and adherence.