The MovieLens Data Set

Michael Ekstrand
GroupLens Research
University of Minnesota

September 13, 2012

Overview

• Three sets of movie rating data (ML-100K, ML-1M, ML-10M)
• Obtained from the MovieLens movie recommender
• Widely used for 10-15 years

MovieLens

Rate movies, get recommendations!

A Little History

• Mid 1990's: DEC EachMovie
• It got shut down, but released data
• We bootstrapped MovieLens with its data
• Started in 1996

MovieLens Stats

• 125K fully-registered users (since 1996)
• 19M ratings of 19K movies
• 2.7K active users in last month

Additional Features

• Buddy Search
• Q&A
• Tags (applications, expressions)
• Movie Tuner
• User-contributed movie data
• Some NetFlix integration

Research

• We do many research projects
• Users know we're a research platform
• We also release data!

MovieLens 100K Data Set (ML-100K)

• 100K ratings, 1-5 stars, timestamped
  • 943 users
  • 1682 movies
  • Light user demographics
  • Movie titles & genre
• Most widely used

MovieLens 1M Data Set (ML-1M)

• 1M ratings, 1-5 stars, timestamped
  • 6040 users
  • 3706 movies
  • Very basic demographics
  • Movie info

MovieLens 10M Data Set (ML-10M)

• 10M ratings, 0.5-5 stars w/ half stars, timestamped
  • 69,878 users
  • 10,677 movies

• Includes 95,580 “tag applications”

  Users can add tags, or thumb-up tags. This records those events.

Licensing

• Free for “noncommercial” use
  • academic, personal good
  • internal R&D in industry fine
  • don't bootstrap commercial system
• Don't redistribute, just send people to us

Useage

No good stats available …

… but over 300 citations

Picking a Data Set

• 100K great for development & testing
• 10M ideal for publishing, but is big
• Don't use 100K for publishing, please

Caveats

• Recommender-influenced (Cosley et al.)
• They are movies
• Users rate 15 moves in initial setup
• Data sets only contain 20+ movie users
• User demographics likely nonstandard

But no data set is perfect. It's just good to understand how each is imperfect.

Thank you