Rate a handful of films and watch three classic approaches—content-based, collaborative, and hybrid—compete in your browser. No server, no API keys: vectors, cosine similarity, and neighbor-based scoring, explained as you go.
Top picks per algorithm (movies you haven’t rated)
Genre vectors + cosine similarity to your taste profile.
Match your ratings to synthetic “neighbor” users, borrow their favorites.
Normalized blend: 55% content + 45% collaborative signal.
Same engine, two worlds—no history vs. enough signal
Visual sketches of what each method is doing under the hood
Each movie is a vector over genres. Your profile is a weighted average of vectors from films you rated. We recommend unseen titles with the highest cosine similarity to that profile—pure item metadata, no other users required.
We compare your ratings to synthetic user profiles that “rated” the same catalog. Neighbors with similar patterns on overlapping movies vote for what you might like next. In production this becomes matrix factorization or deep retrieval at scale.
Content-based systems excel when metadata is rich; collaborative signals capture taste beyond genres. A hybrid smooths blind spots: we min-max normalize both score lists, then combine with fixed weights (here 55% / 45%) and re-rank.