Food recommendation systems have gained increasing attention with the rise of online food delivery and dietary tracking platforms. Unlike traditional recommendation domains, food recommendation must balance user preferences with nutritional health, making the task more complex. Existing systems often optimize for accuracy alone, neglecting the potential health implications of their recommendations. In this paper, we propose a health-aware food recommendation framework that integrates deep representation learning with reinforcement learning. First, a deep sparse autoencoder is employed to extract latent features from three heterogeneous sources: ratings, reviews, and food ingredients. Next, the predicted ratings are refined using a reinforcement learning algorithm that reranks the recommendation list based on a composite objective, combining standard accuracy metric with a health metric. Experimental results demonstrate that our approach not only improves prediction accuracy but also generates recommendation lists that better align with nutritional quality. Our method specifically enhances recommendation performance, improving precision, recall, NDCG, and health metrics by 5.9%, 9.6%, 13.2%, and 30%, respectively, compared to the second-best baseline.
