Similarity beat collaborative filtering by 20%.

A cosine-similarity engine over 5,000+ titles — 20% more accurate than a collaborative-filtering baseline, in my own benchmark.

Pythonscikit-learnPandas

What it does

Feed it a movie. It finds what's actually similar — cosine similarity, not collaborative filtering.

Architecture

Built on scikit-learn and Pandas. The deeper pipeline — feature prep, exact similarity weighting — isn't written up yet. Worth expanding directly rather than guessing at here.

Implementation

Benchmarked against collaborative filtering, came out 20% ahead. The evaluation methodology behind that number is a good candidate for a follow-up pass on this writeup.

Trade-offs

Collaborative filtering was the explicit comparison point. Why the gap landed where it did for this dataset isn't documented yet — said honestly here, not invented.