🎬 Movie Rating Predictor — real IMDb data
Can a network guess a film's IMDb rating before the reviews are in? Pick your inputs — year, duration, popularity, genre — and two engineered features: the director's and lead actor's reputation (their average rating on other films). Shape the network and train (max 50 epochs). Real ratings are noisy, so a good fit is hard-won. Score = test MSE × (1 + 0.4% per parameter), up to 3 attempts.
Data: IMDb Indian Movies (Kaggle) — top 3,500 most-voted rated films. Reputation features are built from the training split only (leave-one-out), so they're fair, not leaky.
FEATURES
What predicts a rating? Not everything does. Click to toggle. 3 selected.
Line blue = positive weight, orange = negative; thickness = strength. Watch them shift as it trains.
OUTPUT
• train • test — perfect fit
🍿 Rate a real movie
🧪 Things to try
One-hot vs reputation
Toggle a few “Directed by …” one-hot inputs, then swap them for Director reputation. One-hot only helps that one director and can't generalize; the single reputation number captures every director. That's why we encode high-cardinality categories instead of one-hot-ing them.
Popularity ≠ quality
Popularity (vote count) nudges the prediction, but blockbusters aren't always well-rated. See how far it gets you alone.
Genre is weak
Genre flags barely move the needle — a hint that some inputs are near-useless and just cost parameters. Drop them to lower your Score.
Real data is humbling
Even your best model leaves lots of scatter (R² well under 100%). You can't fully predict taste from metadata — recognising that ceiling is the lesson.