User-Based Collaborative Filtering for Movie Recommendation: Implementation and Evaluation on the MovieLens Dataset

Abstract

Recommendation systems are an indispensable feature in data mining applications today, owing to their ability to facilitate personal discovery in extensive item sets. In this paper, an end-to-end user-based
collaborative filtering approach is performed using the MovieLens dataset. The implemented pipeline includes validation, preprocessing, analysis, splitting using positive-only leave-one-out sampling, creation
of the user-item matrix, calculation of user-user similarities, rating prediction, generation of Top-K recommendations, and ranking performance. This paper does not introduce any new algorithms; rather, it
provides an experimentally reproducible research on factors influencing the performance of the recommendations generated using user-based collaborative filtering, including evaluation methodology,
sparsity consideration, number of neighbors, rating normalization, and similarity filtering. The results demonstrate that the highest-scoring configuration of raw ratings obtained HR/Recall@10 = 0.355 and
NDCG@10 = 0.216, beating the mean-centered reference configuration while significantly exceeding a random baseline.