MovieLens is a noncommercial movie recommendation website run by the GroupLens research lab at the University of Minnesota, and the name of the family of public rating datasets drawn from it. In Recommender systems research, the MovieLens datasets are the most widely used benchmark for explicit feedback, so the name usually refers to the data rather than to the website itself.
Each dataset records which users rated which movies, with star ratings, timestamps and, in later releases, free-text tags and genre labels. Because the data are real, stable and freely available for research, MovieLens became the common ground on which methods in artificial intelligence for recommendation, from collaborative filtering to neural ranking models, are compared.
How MovieLens collects rating data
MovieLens collects data through an ordinary recommendation service: users sign up, rate movies they have seen on a five-star scale, and receive personalized suggestions in return. The site launched in 1997, shortly after an earlier research service called EachMovie closed, and it has been used continuously since then as a live experimental platform for studies of interfaces, explanations and user behavior.
New members go through an onboarding step in which they rate a set of movies before recommendations appear. This design is the reason most released datasets include only users with at least 20 ratings, which makes the data denser than typical production logs. Since 2003, the interface has accepted half-star increments, so newer datasets use a scale from 0.5 to 5.0 in steps of 0.5, while the oldest releases use whole stars from 1 to 5.
The following table lists the main record types found in MovieLens releases and what researchers typically use them for.
| Record type | Main fields | Typical use |
|---|---|---|
| Ratings | User ID, movie ID, rating, timestamp | Training and testing rating or ranking models |
| Movies | Movie ID, title with year, genres | Content features and genre-aware models |
| Tags | User ID, movie ID, tag text, timestamp | Folksonomy and content-based studies |
| Links | Movie ID mapped to external movie identifiers | Joining metadata such as plots or posters |
| Tag genome | Movie and tag relevance scores from 0 to 1 | Dense item descriptors and explanation research |
The main MovieLens dataset releases
GroupLens has published several fixed snapshots of different sizes, named mostly by their approximate number of ratings. A fixed snapshot never changes, which is what makes results comparable across papers; the “latest” releases, in contrast, are updated over time and are suitable for exploration but not for reproducible benchmarks.
| Release | Ratings | Users | Movies | Notable features |
|---|---|---|---|---|
| ML-100K | About 100,000 | About 940 | About 1,700 | Whole stars, simple demographics |
| ML-1M | About 1 million | About 6,000 | About 3,900 | Age, gender and occupation fields |
| ML-10M | About 10 million | About 70,000 | About 10,700 | Half stars and user tags |
| ML-20M | About 20 million | About 138,000 | About 27,000 | Tags, links and tag genome |
| ML-25M | About 25 million | About 162,000 | About 62,000 | Extended tag genome |
| ML-32M | About 32 million | About 200,000 | About 87,000 | Most recent large snapshot |
Smaller releases such as ML-100K remain popular for teaching and quick experiments, because a model trains in seconds. Larger ones are preferred when a method must show that it scales to millions of interactions, although even ML-32M is small compared with industrial logs that contain billions of events.
Why MovieLens became a standard benchmark
MovieLens became a standard because it combines real user behavior, open access and stable versions, a rare mix in a field where most valuable data belong to companies. Before public datasets existed, each group evaluated on private data, so claims could not be verified. The Netflix Prize dataset later reached a larger audience, but its distribution was withdrawn over privacy concerns, which left MovieLens as the long-term reference point.
The datasets also suit many related tasks in recommendation research. Researchers use them for rating prediction, top-N recommendation, matrix factorization, cold start simulations, sequential recommendation with timestamps, and studies of popularity bias and diversity. The table compares MovieLens with other public data commonly used in the same studies.
| Dataset | Feedback type | Domain | Main strength | Main weakness |
|---|---|---|---|---|
| MovieLens | Explicit star ratings | Movies | Stable, clean, well documented | Dense and unlike production data |
| Netflix Prize data | Explicit star ratings | Movies | Very large for its time | No longer officially distributed |
| Amazon review data | Ratings and review text | Retail products | Many categories and text | Extremely sparse, noisy duplicates |
| Click and session logs | Implicit feedback | E-commerce or news | Close to real ranking tasks | Often short-lived or restricted |
Using MovieLens correctly in experiments
Correct use of MovieLens starts with naming the exact release and the evaluation protocol, because results on ML-1M and ML-20M are not interchangeable. A reported score is meaningful only together with the data split, the metric and the candidate set used for ranking.
Many modern papers convert explicit ratings into implicit feedback, for example by treating every rated movie as a positive interaction or keeping only ratings of four stars and above. This conversion is legitimate, but it changes the task from predicting stars to predicting which movies a user chose to rate, and that difference must be stated. The decision table summarizes the most common protocol choices.
| Choice | Recommended option | Reason |
|---|---|---|
| Data split | Time-based or leave-last-out per user | Random splits leak future information |
| Rating task metric | RMSE or MAE | Measures error on the star scale |
| Ranking task metric | Recall, NDCG or MAP at a fixed cutoff | Matches how lists are shown to users |
| Candidate set | Full catalog ranking | Sampled negatives can reverse model rankings |
| Baselines | Tuned matrix factorization and popularity | Weak baselines inflate reported gains |
The data are released under terms that allow research use, require acknowledgment of GroupLens, and prohibit commercial use or redistribution without permission. Teams building products therefore use MovieLens for prototyping and method selection, not as training data for a commercial service.
Limitations and common mistakes with MovieLens
The biggest limitation is that MovieLens is not representative of production traffic. Its users are self-selected movie enthusiasts, every user has many ratings, and ratings were often entered in bursts during onboarding rather than after watching, so timestamps do not reflect viewing order.
- Overreliance on one dataset: a method that wins on MovieLens may lose on sparse retail or news data.
- Timestamp misreading: sequential models can learn rating sessions instead of real taste evolution.
- Sampled evaluation: ranking the true item against 100 random negatives tends to overstate accuracy and can change which model looks best.
- Version confusion: mixing results from different releases makes comparisons invalid.
- Ignoring popularity effects: a simple most-popular list is a surprisingly strong baseline on these data.
Understanding it simply
MovieLens is a website where people rate films and get suggestions for what to watch next. The researchers who run it regularly publish the collected ratings, with names removed, so that anyone can test new recommendation ideas on the same real data.
Think of it as a shared practice exam for recommendation engines. Because everyone answers the same questions, it is easy to see whose method does better. The catch is that a practice exam is cleaner than real life, so a good MovieLens score is a promising sign, not a guarantee.
Applied in practice
A small data science team at a regional video streaming service wants to replace its genre-based suggestions with a learned model. Before touching its own logs, the team downloads ML-25M and compares a popularity baseline, alternating least squares and a two-tower neural model under a time-based split, measuring recall at 20 over the full catalog.
On MovieLens, the neural model beats matrix factorization by a clear margin, while both easily beat popularity. When the team repeats the same pipeline on three months of its own viewing data, the gap between the two learned models shrinks to almost nothing, because its catalog is smaller and most users watch only a few titles. The outcome is a faster, cheaper launch with matrix factorization, and MovieLens served its proper role: a safe place to build and debug the pipeline before the decisive test on real traffic.