Comparative Study of Memory-Based and Model-Based Collaborative Filtering for Food Recommendations

Food recommender systems are essential for providing consumers with tailored recipe recommendations according to their dietary requirements and preferences. Traditional recommendation approaches, such as collaborative filtering, have been widely used but require a thorough evaluation to determine their effectiveness in the food domain. The Allrecipes.com and Food.com datasets were used in this study to compare and contrast memory-based and model-based collaborative filtering for food suggestion. Using Cosine, Pearson, Jaccard, and Euclidean similarity metrics, the memory-based method uses item-based and user-based collaborative filtering.We evaluate several model-based techniques—Deep Matrix Factorization (DeepMF), Non-Negative Matrix Factorization (NMF), Singular Value Decomposition (SVD), and its enhanced variant Singular Value Decomposition++ (SVD++). Their effectiveness is measured using F1-score, Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE). The results demonstrate that model-based approaches consistently exceed memory-based methods in accuracy, with DeepMF delivering the best performance. Among similarity measures, Cosine similarity gives the best result for memory-based filtering. The results point towards the requirement of proper recommendation methods and show potential future improvement with hybrid models as well as enhance explainability of food recommender systems.