Information

Authors: James McInerney, Benjamin Lacker, Samantha Hansen, Karl Higley, Hugues Bouchard, Alois Gruson, Rishabh Mehrotra
Date: September 2018
Published By: RecSys '18: Proceedings of the 12th ACM Conference on Recommender Systems

Link

Although, in existing works, bandits and explanations are separated, this study treated both methods at the same time.
- 1. learn which explanations each user responds to.
- 1. learn the best content to recommend for each user.
- 1. balance exploration with exploitation to deal with uncertainty.
Proposed "Bart (Bandits for Recsplanations as Treatments)" based on a contextual bandits that addresses the problem of recommending with explanations under uncertainty of user satisfaction (problem: How to incorporate exploration when jointly optimizing for items and explanations. Naively treating each (item, explanation) pair as a distinct action would multiplicatively scale the number of actions that need to be explored.)
Through offline experiments on randomized historical data and online experiments with users in the homepage of a large-scale music recommender system, it is empirically evaluated and found that Bart significantly outperforms the best static ordering of explanations.

*Recsplanations (a term coined in Spotify): to tell a user why they are being recommended a particular item.