MovieLens Datasets: Context and History

By on

The MovieLens datasets are full of data describing how people rate movies. As it turns out, these datasets have been useful to lots of folks, from recommender systems researchers to the readers of popular-press programming books. Though it is difficult to measure the full extent of the datasets’ impact, we see that they were downloaded more than 140,000 times in 2014, and that the keyword “movielens” currently results in over 8,900 results in Google Scholar.

It is tempting to view these collections of ratings as a cohesive whole. However, the truth of the matter is that the datasets are the product of 17 years of member activity in a web site that has seen its fair share of changes and experimental features. Given the extent of attention — research and otherwise — given to these datasets, it seems worth exploring the relationship between the system and the resulting data.

(more…)

Putting Users in Control of their Recommendations

By on

The music services that I subscribe to don’t understand me very well. Pandora, which puts together personalized radio stations, seems to think that I only like the very most popular music, which I don’t. Spotify, which offers a new personalized playlist for me each week, seems to think that I only like quite obscure music. But neither of them get it right, and I wish that I could tell them to change.

screenshots of pandora and spotify
Pandora and Spotify are “black box” recommenders, where it is difficult to know how to act to repair bad recommendations.

(more…)

GroupLens has gathered for a photograph

By on

Look at this group of nice folks! This GroupLens group photo was taken in the atrium of Keller Hall, where we work.

grouplens-2013

GroupLens, Fall 2013, back to front:

  • Jacob, Raghav, Zihong, Kenny
  • Yilin, Brent, Morten
  • Pik-Mai, Steven, Dan
  • Zihong, Derian, Fernando
  • Vlad, Anu, Michael
  • Ting, Alison, Loren
  • Vikas, Max, Joe
  • Andrew, Daniel, Tien