Context-based User Playlist Analysis for Music Recommendation

Abstract

Convenient access to music through streaming platforms has given rise to an insurmountable amount of choice when it comes to listening to music. These platforms have turned to music recommender systems to keep the user engaged by giving personalized recommendations. In recent years these algorithms have made great strides and seen huge improvement. However, these music recommender systems can enforce certain biases and cause a lack of diversity within their recommendations. Research has focused on countering these problems with the use of context-dependent recommender systems. Interestingly, there has been a lack of focus on activity based music listening behavior. This study uses different analysis methods to research the correlation between user activity context and musical preferences. Results show that there are significant differences between different activities and the musical features that are contained within a song. Thereby suggesting a use for activity context within music recommender systems. Contrastingly, results from the clustering, classification and the user survey show that it remains difficult to determine which songs are listened to in which contexts of activity. On top of showing that musical taste can not solely be determined by activity, these results show that musical preference remains distinctly subjective and recommendation algorithms will forever struggle in determining the right music for the right person at the right time. Concluding, while activity context shows promise in being useful in recommending music and helping overcome biases and lack of diversity within recommendations, an activity based method should be combined with other algorithms such as content based recommenders. Thereby helping to adhere to users’ broad and expansive musical preferences while ensuring relevant and personal recommendations.