Yum-Me: A Personalized Nutrient-Based Meal Recommender System.

Area of Science:

• Artificial Intelligence
• Nutrition Science
• Human-Computer Interaction

Background:

• Nutrient-based meal recommendations can aid in managing diabetes and obesity.
• Current systems struggle with accurately capturing fine-grained food preferences and nutritional needs simultaneously.
• Existing preference learning methods are often high-level or require extensive user input over time.

Purpose of the Study:

• To introduce Yum-me, a personalized nutrient-based meal recommender system.
• To develop and evaluate advanced methods for food preference learning and food image analysis.
• To improve the accuracy and efficiency of personalized meal recommendations.

Main Methods:

• Developed Yum-me, a system integrating a visual quiz-based interface for preference profiling.
• Introduced FoodDist, an open-source food image analysis model.
• Implemented a novel online learning framework for preference learning from image comparisons.
• Conducted a 60-person user study for end-to-end validation.

Main Results:

• FoodDist demonstrated superior performance in food image analysis.
• The online learning framework significantly outperformed baseline methods in a field study with 227 users.
• Yum-me improved the meal recommendation acceptance rate by 42.63% in a user study.

Conclusions:

• Yum-me effectively addresses challenges in personalized nutrient-based meal recommendations.
• The developed food image analysis and preference learning techniques offer significant advancements.
• Yum-me shows strong feasibility and effectiveness in improving user engagement with dietary recommendations.