Finding Groups

Movie Recommender

Twenty kids have rated these made-up movies. Rate a few yourself and the robot will find the kids whose taste looks like yours, then guess which movies you'll love.

Tap the stars to rate 4 or 5 movies you think you would like or hate. I'll take it from there!

Rate some movies

Robo's top picks for you

Rate a few movies first.

Your taste twins

Challenges (they get a numbered flag on the chart)

1 The cheater: give every movie 5 stars. Can the robot learn anything about you?
2 A brand-new movie that nobody has rated. What can the robot say about it?
3 Odd taste: like and dislike movies in a way that matches nobody (best twin below 35%).
4 Fewer ratings: let the crowd keep only 30% or less of their ratings.

How it works

You rate

Your stars are your taste fingerprint.

Find twins

Robo checks which kids gave similar stars on the same movies.

Ask the twins

For a movie you haven't seen, Robo listens most to your closest twins.

Test itself

Robo hides some known ratings, guesses them, and checks how far off it was.

What just happened?

For grown-ups: what's really going on

This is collaborative filtering: recommending from patterns in a user-item rating matrix with many missing entries, with no information about the movies themselves. The 20 viewers and 12 movies are synthetic: three hidden taste groups plus noise, with about a third of ratings missing.

  • User-based CF: similarity between two viewers is computed on movies both rated. "Compare tastes" is the Pearson / mean-centred cosine similarity (each person's average is subtracted first); "Compare raw stars" is plain cosine similarity, which is high for almost everyone because ratings are all positive. The prediction is your mean plus the similarity-weighted average of each neighbour's deviation from their own mean, using the top-k neighbours. Similarities with few shared movies are shrunk.
  • Matrix factorisation: rating ≈ μ + bᵤ + bᵢ + pᵤ·qᵢ with rank 2, trained by stochastic gradient descent with L2 regularisation. The 2-D vectors are latent factors, which is why the taste map can be drawn. The axes themselves have no fixed meaning. Both charts use hold-out ratings the model never trained on (RMSE); the training error keeps falling while hold-out error can stall or rise, which is overfitting.
  • Cheating / no signal: constant ratings have zero variance after centring, so the similarity is undefined and the model falls back to the crowd's average, a popularity baseline.
  • Cold start: an item or user with no ratings has no similarity to use. Real systems fall back to popularity or content features (genre, text) and hybrid models.
  • Sparsity: fewer observed ratings means fewer co-rated items, noisier similarities and worse error.
  • Filter bubbles: recommending only what your twins liked narrows what you see, and what is shown shapes what gets rated next (a feedback loop). Real systems add exploration and diversity.

Further reading: Koren, Bell and Volinsky, "Matrix Factorization Techniques for Recommender Systems" (2009) · the Netflix Prize.