MovieLens 20M movie ratings. Stable benchmark dataset. 20 million ratings and 465,000 tag applications applied to 27,000 movies by 138,000 users. Includes tag genome data with 12 million relevance scores across 1,100 tags. Released 4/2015; updated 10/2016 to update links.csv and add tag genome data.
Use case
Optimize over a session the recommended film (or series) in order to maximize viewer satisfaction and retention
| Example observations | user-id (anonymised) Aggregate profile: mean and variance of past ratings (1-5), preferred top-N genres, last film seen, time since the last interaction Time context: timestamp (day-week/evening vs. morning) derived from timestamp Features of the film proposed at the previous step (gender, year of release) (Indicative dimension: 1 id, 10-30 digital features plus 20 binary types) |
| Example actions | movie_id (2,278 tracks) (In a bandit implementation of the actions are the film identifiers available at this moment; in a sequential approach, the “candidate set” can be restricted (e.g. the 1,000 most popular films or filter preselected).) |
| Example reward | (rating – 3)/2 → scale∈ [-1, +1] Variant: bonus of + 0.1 if the user leaves another note in the same session (retention proxy) |