Title: Recommender Systems Unplugged: Effects of explaining algorithmic recommendations on music discovery, an experimental approach 

Abstract:

Music streaming services make massive use of algorithms in their music recommender systems (MRS) to guide users to tracks that they are likely to enjoy. However, the black-box nature of these algorithms makes them difficult for users to understand, both in terms of how they work and the music they predict. The field of explainable AI (XAI), and in particular its “explanation” side, has emerged to make the uses of AI (including MRSs) more comprehensible to users. This paper aims to observe, using an experimental method, whether the explanation of an MRS algorithm induces a change in discovery behavior on music streaming services. In a theoretical framework, we model two types of user discovery behavior, namely “study” and “browse” behaviors. We then test in the lab the explanation effects on these behaviors by explaining a simplified “semi-personalized” MRS, and measuring the relative listening time of the tracks. We observe no average effect induced by the explanations, but we observe a differentiated impact of explanations based on the treatment intensity (i.e. the time spent reading to it), the more the people are treated, the more they listen to the tracks, reinforcing the “study” behavior. 

Title: Recommender Systems Unplugged: Effects of explaining algorithmic recommendations on music discovery, an experimental approach 

(co-writing with Julien M'Barki)

Abstract:

Music streaming services make massive use of algorithms in their music recommender systems (MRS) to guide users to tracks they are likely to enjoy. However, the black-box nature of these algorithms makes them difficult for users to understand, both in terms of how they work and the music they predict. The field of explainable AI (XAI), and in particular its “explanation” side, has emerged to make the uses of AI (including MRSs) more comprehensible to users. This paper aims to observe, using an experimental method, whether the explanation of an MRS algorithm induces a change in discovery behavior on music streaming services. In a theoretical framework, we model two types of users' discovery behaviors, namely “study” and “browse” behaviors. We then test in the lab the explanation effects on these behaviors by explaining two simplified MRSs, taking into account only certain music recommendation criteria.