Personalized Interfaces for Music Recommender Systems

Recommender systems are software tools that automatically select a set of relevant items for users based on their interests. Despite their intelligence to adapt the recommended items to the user, most recommender systems lack the intelligence to adapt their interface to the needs of the user. Additionally, most recommender systems do not allow users to steer the recommendation process or correct wrong assumptions. Moreover, recommender systems often act as black boxes which might hinder users to steer the recommendation process in an informed way. In this dissertation, we take the first steps towards personalized recommender systems that can adapt themselves based on the personal characteristics of the user. In particular, we investigate how the system can adapt visualizations to control the recommendations and how it can adapt scrutable explanations that allow both control and understanding of the recommender system. More concretely, this dissertation aims to address three research questions: \\ RQ1 - How do personal characteristics influence the way users perceive and interact with different visualizations to control a music recommender system?\\ RQ2 - Which personal characteristics influence the way users perceive and interact with scrutable explanations in a music recommender system? \\ RQ3 - How should explanations be adapted to different personal characteristics in a music recommender system. To investigate these research questions, we conducted eight different user studies in which we investigated how users interacted and perceived a music recommender system interface. In particular, we focused on the reaction of users to visualizations to control the recommender system and on scrutable explanations. Additionally, we observed how personal characteristics such as personality, cognitive style, cognitive abilities, and domain experience influenced the reaction of users to these interfaces. Our results indicate that two personal characteristics influence the way users perceive and interact with visualizations to control recommender systems, namely musical sophistication, and tech-savviness. We also found three personal characteristics that influence the perception of scrutable explanations in a music recommender interface need for cognition, musical sophistication, and openness. In three user studies, we investigated how scrutable explanations could be tailored to these three personal characteristics. Based on our results we recommend providing explanations up-front for all recommendations at once, but for low NFC users it must be possible to turn off these explanations. To tailor explanations to musical sophistication, we recommend providing brief explanations which do not require musical knowledge to users with low musical sophistication. For users with high musical sophistication, we recommend providing these users with the choice between these brief explanations and interactive explanations which contain a mix of information. To adapt explanations to low openness, we recommend providing explanations with only one explanation element. For high openness, we recommend providing users the choice between these explanations and explanations that support exploration and that contain multiple explanation elements.

Publication year:2021

Accessibility:Closed