Comparison of Collaborative Filtering and Content-Based Filtering for Recommendation Systems

Abstract

The objective of food recommender systems is to provide recipe recommendations that are both personally relevant and technically accurate. This thesis assesses the predictive accuracy and user acceptance of two algorithmic approaches: Content-Based Filtering (CBF) and Collaborative Filtering (CF). An online survey study with 100 participants (from India and the Netherlands; 132 initially recruited) was conducted to measure outcomes such as perceived relevance, satisfaction, trust, and intention to reuse, while offline experiments assessed predictive error (MAE, RMSE). A mixed-method evaluation was conducted. The results indicate that CF achieved a lower predictive error than CBF in the offline evaluation. Acceptance of the system was inconsistent in the user study; 35% of participants indicated that they were highly likely to reuse it, 30% that they were very likely to do so, and 12% that they were unlikely to do so. Subgroup analyses revealed that user characteristics influenced perceptions. For instance, participants with stricter dietary preferences (e.g., vegetarian/gluten-free personas) reported higher satisfaction with CBF, whereas broader user groups preferred CF for its novelty and relevance. These results indicate that no single algorithm is the most effective in terms of all criteria. In general, CF exhibits superior predictive performance, whereas CBF provides benefits in terms of transparency and dietary alignment. This thesis emphasizes the complementary strengths of CBF and CF. It offers insights for the development of more effective and reliable food recommender systems by integrating algorithmic evaluation with user-centered feedback.