Contextual Personalized Re-Ranking of Music Recommendations through Audio Features Based User Preference Models

Abstract

With advancements in Internet and technology, it has become increasingly easy for people to enjoy music. Users are able to access millions of songs through music streaming services like Spotify, Pandora, and Deezer. Access to such large catalogs created a need for relevant song recommendations. Music recommender systems assist users in finding the most relevant songs by consistently matching them with the user’s preference. Accurately representing these preferences is essential to creating accurate and effective song recommendations. User preferences are highly subjective in nature and change according to context (e.g., music that is suitable for running is not suitable for relaxing). Preferences for songs can be based on characteristics of high level audio features, such as tempo and valence.
This thesis proposes a new contextual re-ranking algorithm, which belongs to the group of contextual post-filtering techniques, to leverage users’ contextual information. The algorithm uses two models, a global and personalized model, to model user preferences. These models use audio features to represent user preference in specific contextual conditions. The algorithm is able to re-rank any given music recommendation list. First, we analyze the correlation between audio features and contextual conditions. This analysis shows that the correlations are significant, thus audio features are suitable for representing user preference in contextual conditions. Thereafter, we implement and evaluate the re-ranking algorithm using accuracy metrics on the #NowPlaying-RS and InCarMusic datasets, using various initial recommender algorithms. Results show there
is merit in applying such a re-ranking algorithm to increase recommendation accuracy. The personalized model, given enough historical data, consistently outperforms the global model.