2009:Music Recommendation Personalized Radio

Personalized Radio Subtask

Description

This task is to take a user profile, and without additional information, generate a playlist for this user.

Personalized Playlist Generation With Social Data

• Source Data
  • Audio content
  • User profiles
  • User playcounts of audio content
  • Song tag lists
  • artist tag lists

• Query Data

User id

• Ground Truth

Original playcount data on user id's

• Execution

Perform cross-validation on user playcounts, providing all audio and associated data as additional data

• Evaluation

For those algorithms outputing a numeric output (strength of a recommendation)

• • Mean Error Recommendation
  • RecommendationError

For those algorithms outputing ordered sets

• • ROC Area
  • Kendall Tau

For those algorithms outputing unordered sets

• • Pearson Correlation

Playlist Generation With Time Dependant Social Data

Same as above, but the ground truth is the difference between consecutive playcounts and the cross-validation is performed on the ground truth

Personalized Playlist Generation of New Music

Same as untiumed data, but the social data is witheld on a percentage subset, simulating new music. For user playcounts, this means removing these songs from the source data. The evaluation is cross-validation of those user playcounts that contain music with data witheld, recreating the removed entries for new music without providing any of the removed playcounts during training.