Palate

Palate predicts the rating you would give a film, rather than the rating the average viewer gave it. It runs over two real datasets — 1.40 million Rotten Tomatoes critic ratings and 11.08 million Letterboxd user ratings, 12.48 million in total — with explicit train/held-out splits.

Three models are benchmarked against a plain mean-score baseline: an interpretable similarity-based analytic model, gradient-boosted trees (XGBoost), and a small PyTorch neural network. A collaborative-filtering step picks the nearest-neighbour films for a user to rate, which improves the input signal the prediction models get. An error attribution showed most of the gain comes from calibrating the score scale, with taste matching contributing less at this sample size.

Try the web app →

RMSE versus number of user ratings for three models — Held-out RMSE against n, the number of ratings a user has already supplied (Rotten Tomatoes design; shaded bands are the spread over resampled profiles). XGBoost and the neural network are already better than the crowd average at n = 3, and stay close to each other throughout. The analytic similarity model starts far worse but improves steeply, passing the baseline somewhere in the low tens of ratings.