MovieLens
GroupLens Research, University of Minnesota
32 million movie ratings from 200,000 people: the benchmark behind decades of recommender research.
Ratings and tags collected from users of the MovieLens movie recommendation site, released in fixed versions from 100K (1998) to 32M (2024), plus Tag Genome relevance scores. Each rating row is one user's 0.5-5 star score for one movie with a timestamp. It is the standard benchmark for recommender systems and a ready source for studying taste.
Access
- Cost
- FreeFree; GroupLens asks users to read README terms and submit a request form, and does not allow public redistribution.
- License
- Open, non-commercial onlyGroupLens usage terms in each README (research use, no redistribution)
- Sign-up
- Application required
- How you get it
- Bulk downloadCSV, NPZ (1B synthetic)
What's in it
- One row is
- One row per user-movie rating (also tag applications and movie metadata)
- Covers
- MovieLens users (mostly US)
- Years
- 1995-2023 (ratings in the 32M release)
- Updated
- Irregular
- Size
- 32 million ratings from 200,948 users on 87,585 movies (32M release)
Quality
Stable, versioned, documented releases used in thousands of papers.
- Users are self-selected recommender-site users, not a random sample
- Latest releases change over time; GroupLens says not to use them for reported results
Connects to
- IMDb Non-Commercial Datasets: Links file maps movies to IMDb IDs for credits and ratings
- TMDB API: Links file maps movies to TMDB IDs for budgets and posters
- UCSD Goodreads Datasets: A parallel ratings data set for books
Questions it could answer
- Do older movies get higher average ratings, and is that survivorship?
- How do individual users' ratings drift over time?