Population-Level Analysis of Personalized Food Recommendation Using Reinforcement Learning

Area of Science:

• Food industry innovation
• Consumer behavior analysis
• Recommender systems

Background:

• Traditional recommendation systems often neglect diverse consumer preferences.
• Understanding cultural and age-based nuances is key to enhancing consumer appeal and marketing strategies.
• Personalized recommendations can optimize food industry operations.

Purpose of the Study:

• To introduce a novel methodology for optimizing recommendation strategies in the food industry.
• To evaluate the performance of different recommendation algorithms across simulated diverse populations.
• To demonstrate how understanding population-specific preferences can drive marketing opportunities and operational efficiency.

Main Methods:

• Simulated diverse populations using fuzzy logic based on characteristics like age, gender, and location.
• Evaluated recommendation algorithms including State-Action-Reward-State-Action (SARSA), multi-armed bandit (MAB), and Deep-Q Network (DQN).
• Assessed algorithm performance within a generated menu database, measuring accumulated reward and statistical significance.

Main Results:

• The Deep-Q Network (DQN) significantly improved accumulated reward across various simulated populations (e.g., 71.60% for "Foodies").
• Multi-armed bandit (MAB) offered comparable performance to DQN with greater resource efficiency.
• Statistically significant performance differences were observed for DQN across populations (p < 0.005).

Conclusions:

• Recommender systems offer a strategic advantage for navigating market demand, optimizing supply chains, and reducing food waste.
• Tailoring recommendations to specific demographics, like "Foodies" or "Seniors," enhances consumer engagement.
• The optimal recommendation strategy requires balancing algorithmic effectiveness, computational cost, and industry-specific needs.