The problem: recommend films someone is likely to enjoy next, using only what's known about the films: what they're about and who made them. No viewing history is needed, so it works for new users and new titles.
What I built: a recommender over 28,932 films from the TMDB April 2024 export. It blends plot similarity with director and cast similarity, and I tested it fairly against random and most-popular baselines.
Results
Average genre overlap between each film and its 10 recommendations, on 2,000 test films.
| Model | Share genre | Genre overlap |
|---|---|---|
| Random films | 43.9% | 0.154 |
| Most popular films | 36.9% | 0.103 |
| People only (director + cast) | 65.4% | 0.298 |
| Plot only (TF-IDF) | 64.2% | 0.284 |
| Hybrid | 71.0% | 0.333 |
The hybrid more than doubles random genre overlap and beats either signal on its own. Its picks have the catalogue's typical popularity, so it isn't just pushing blockbusters.
Because you watched Spirited Away: The Boy and the Heron, Princess Mononoke, Ponyo, My Neighbor Totoro, Howl's Moving Castle.
How I tested it
- Genres were never given to any model. They were held back as the answer key, so the score can't be gamed.
- The blend was chosen on 1,000 validation films (60% plot, 40% people), then every model was scored once on 2,000 different films.
- Two baselines: random films, and the most popular films. The most popular films score below random, which shows why popularity alone makes a poor recommender.
How it works
- Plot: TF-IDF on each overview, with unigrams and bigrams and 60,000 features
- People: each director and each of the top 5 billed cast is one token, weighted so prolific actors count for less
- Similarity search across the full catalogue, in batches so it runs on a laptop with no GPU
- Data step: 907,610 titles streamed from a 292 MB Excel file, keeping films with more than 50 votes
Limitations
- Genre agreement is a proxy for relevance. With viewing histories, I'd measure precision on held-out watches.
- The director signal can dominate. The Dark Knight pulls in Christopher Nolan's short films, so down-weighting shorts is the next fix.
This is a rebuild of an earlier version whose evaluation was a placeholder. The new version replaces it with the held-out test above.