CineMatch · Movie Recommender

The problem

Modern platforms drown users in content, and the typical student recommender stops at 'it returns movies.' Our team wanted one that could answer a harder question: which collaborative-filtering approach is actually better on this data, and by how much?

Approach

• Built a FastAPI + SQLite backend serving top-K neighbor predictions over a normalized schema (users, movies, ratings, tags, links).
• Implemented both user-based and item-based collaborative filtering with Pearson and cosine similarity over mean-centered ratings.
• Cleaned the data the way it actually needs cleaning: dedupe, keep-latest on repeat ratings, filter low-activity users/items, parse the release year out of titles, normalize genres, and a time-aware train/test split so the test set represents future behavior.
• Added a baseline correction (global mean + user bias + item bias) and evaluated with RMSE, MAE, Precision@K, and coverage, measured, not eyeballed.
• Built the frontend in React + Vite + Tailwind against the REST API.

Impact

• Side-by-side, measured comparison on a held-out time-aware split: item-based hit RMSE ≈ 0.891 / MAE ≈ 0.666 at 81.5% coverage, user-based hit RMSE ≈ 0.905 / MAE ≈ 0.689 at 90.8% coverage, the accuracy-vs-coverage trade-off you only see if you actually measure it.
• Clean separation: typed REST API, normalized SQLite schema, reproducible preprocessing pipeline.

Stack