<![CDATA[The growing volume of content on film streaming platforms makes it harder for users to find films that suit their tastes, highlighting the need for a personalized recommendation system. This research introduces a Hybrid Filtering architecture that unifies Collaborative Filtering using Singular Value Decomposition with Content-Based Filtering employing TF-IDF weighted vectors and cosine similarity measurement. The defining innovation is a dual-threshold mechanism grounded in each user's personal average rating (user_avg_rating), which individually filters candidate films before their hybrid scores are combined. New users are handled through a genre-based onboarding flow supported by a weighted popularity score. Experiments on the MovieLens ml-latest-small dataset (100,836 ratings, 610 users, 9,742 films) show that Hybrid Filtering achieves the best Precision@10 of 0.27 and NDCG@10 of 0.32. Switching from a fixed global threshold to the personal threshold raised Genre Precision from 0.40 to 0.90 (+125%). The SVD component attained RMSE of 0.88 and MAE of 0.67, and a user study with 10 participants yielded a mean relevance score of 8.2 out of 10. Keywords—Hybrid Filtering, Collaborative Filtering, Content-Based Filtering, Personal Threshold]]>
Copyrights © 2026
