29th Conference on User Modeling, Adaptation and Personalization

Online from Utrecht, the Netherlands

June 21-25, 2021

joomla social media module

acm

sigchi

sigweb

universiteit utrecht

radboud universiteit

Target-aware Aggregate Diversification in Recommendation

Himan Abdollahpouri (Northwestern University), Jesse Anderton (Spotify), Zahra Nazari (Spotify), Ben Carterette (Spotify)

 

Recommender systems with different groups of items

  • The recommendation platform defines a fair outcome in terms of what percentage of recommendations should be allocated to what item group
  • It is important to ensure minimum deviation from the users' preferences

 

Example of three target distributions for the exposure of different groups of items

 

Popularity Fairness

  • Making sure items from different popularity groups (Head, Mid, and Tail) are fairly represented
  • The system defines what is fair

users propensity towards item popularity

 

Ratio of different item popularity groups in recommendations:Ratio of item groups in recommendation list

Ratio of different item popularity groups in user's profile:

Ratio of different item popularity groups in user's profile

Method 1 (Target): maximizing target calibration without considering users' tolerance:

Target calibration

Method 2 (Target+U): maximizing target calibration and user calibration simultaneously:

Target and user calibration 

 

Evaluation Metrics

  • Precision
  • Target Miscalibration
  • User Miscalibration
  • Coverage

Experiment Settings

  • Data: MovieLens 1M
  • Three item popularity groups: Head, Mid, and Tail
  • Two target distributions: [0.2,0.6,0.2], [0,0.5,0.5]

Results for Target= [0.2,0.6,0.2]

Precision, target calibration, user calibration, and item coverage for target [0.2,0.6,0.2]

popularity in data vs popularity in recommendations

Ratio of item groups in recommendation list for target=[0.2,0.6,0.2]

 

 

Results for Target= [0,0.5,0.5]

Precision, target calibration, user calibration, and item coverage for target [0,0.5,0.5]

popularity in data vs popularity in recommendations

 

Ratio of item groups in recommendation list for target=[0,0.5,0.5]

Questions:

  1. What are some other ways to incorporate the system's target distribution into account ?
  2. How can the defined target distribution be modified such that the final recommendations still are close enough to that target yet the user satisfaction is less affected?