A Hybrid Context-Aware Approach For Personalized Food Recommendation

Abstract As food delivery platforms are rapidly growing, so are the needs to use personalized food recommendation systems that will make the experience of the user more enjoyable. Rising food delivery platforms wishing to boost the user experience and user engagement have put extreme demand for personalized food recommendation systems. This paper proposes the architectures, implementation and comparative study of four food recommendation models: Content-based Filtering, Singular Value Decomposition (SVD++) Collaborative Filtering, Time-Season-Aware-Meal-Temporal and Item-Feature User Preference Cycle(Time-Season-Aware-Meal-TIFU-CYCLE) and a Context-Aware Hybrid Model. On the other hand, Content-Based approach utilize Term Frequency–Inverse Document Frequency (TF-IDF) vectorization to encode item features including food type, product name and outlet features separately, calculates a similarity score based on content between User items that interacted with and the potential items for recommendation. The Collaborative Filtering model uses SVD + + to model latent user preferences based on past user–item interactions. A Time-Season-Aware-Meal-TIFU-CYCLE Model is implemented to account for food reorder cycles, meal-time preferences, seasonal preferences and recency. Content-Based and Collaborative signals are not sufficient to overcome the limitations of each approach, so the Context-Aware Hybrid Model uses both types of signals and contextual parameters such as meal time, season, location etc. The models were tested with Precision, Recall, F1-Score, Hit Rate, NDCG and MAP. The Context-Aware Hybrid Model achieved optimal output with 51.16% Precision@5, 82.35% Recall@5, 52.56% F1@5, 100.00% Hit Rate@5, 95.27% NDCG@5 and 93.37% MAP@5.