Item-Based Collaborative Filtering with Cosine Similarity for Personalized Movie Recommendation: A Comparative Study on the Movielens 100K Dataset

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommendation systems help users find useful items in very large catalogues. Collaborative filtering (CF) is one of the most common approaches, but its results depend on the similarity measure, the neighbourhood size and the way scores are turned into a ranked list. This paper compares user-based and item-based CF with cosine similarity on the MovieLens 100K dataset (943 users, 1,682 movies, 100,000 ratings), using the five official train/test splits. We also include simple baselines and a biased matrix factorization (MF) model. Two tasks are studied: rating prediction (RMSE, MAE) and top-10 recommendation (Precision@10, Recall@10). Neighbourhood size and a shrinkage term were chosen on validation data taken from the training set. MF gave the lowest rating error (RMSE 0.926). User-based CF (0.942) and item-based CF with adjusted cosine (0.947) were close, and their difference is smaller than the fold-to-fold variation. For top-10 recommendation, ranking by a sum of similarities gave a Precision@10 of 0.267 (item-based) and 0.269 (user-based), against 0.157 for a popularity baseline, while ranking neighbourhood methods by predicted rating gave a Precision@10 near zero. On this small dataset we found no evidence that item-based CF is more accurate than user-based CF