A hybrid movie recommendation engine combining collaborative filtering (SVD) with content-based genre similarity. Search movies you like, get personalized recommendations.
Most recommendation systems need extensive user history to work well. MatchMaker solves the cold-start problem by letting users pick movies they like, then combining collaborative filtering (what similar users liked) with content-based similarity (same genres).
Uses scikit-surprise SVD matrix factorization trained on 100K user-movie ratings. Predicts how much a user would like unseen movies based on patterns from similar users.
Builds genre profiles using TF-IDF vectorization and cosine similarity. When you like 'Toy Story', it finds other animation/children/comedy movies regardless of user ratings.
Combines SVD and content-based scores with a configurable weight (60% collaborative, 40% content-based). Small random jitter ensures results vary between requests.
Deployed as part of the najdiavto-ml service on Railway. Four endpoints serving search, recommendations, similar movies, and popular picks.
Uses MovieLens latest-small (9,742 movies, 100K ratings, up to 2018) with 38 injected modern films up to 2024 for complete coverage.