Evaluating Deep Learning and Traditional Approaches in Recommender Systems
Kitsanachai Kairassamee, Karn Yongsiriwit · 2024
Currently, Recommendation systems (RS) have a major role in the selection and presentation of content related to the needs and interests of users. Because of this, we have benchmarked five models: Bayesian Personalized Ranking (BPR), Item-based k-Nearest Neighbors (ItemKNN), Item Popularity (ItemPop), Neural Collaborative Filtering (NCF), and TensorFlow Recommenders (TFRS). Using the Hit Ratio (HR) and Normalized Discounted Cumulative Gain (NDCG) metrics, our benchmark experiment, conducted on the MovieLens 1M and Goodreads datasets, finds the TFRS model to demonstrate superior accuracy and ranking efficiency compared to other models. TFRS consistently achieved the highest HR@K and NDCG@K scores, making it the most effective in capturing the complex relationships between users and items. While the deep learning models capture user-item interactions more effectively of the user-item interaction—BPR, NCF, and TFRS— their counterparts, non-deep learning models—ItemKNN, ItemPop—are faster but less effective in personalization. These insights guide the selection of the most suitable system for enhancing user engagement.