CONTENT-BASED MULTI-CONTEXT-AWARE MUSIC RECOMMENDER SYSTEM DEVELOPMENT WITH BAYESIAN NETWORK

• Main Author: Richard, Ferdinandus
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/39853
• Institution: Institut Teknologi Bandung

Recent few music recommender systems are still using conventional recommender system algorithms, which are content-based filtering and collaborative filtering. However, these algorithms sometimes could still give recommendation that doesn’t fit user’s preferences. Therefore, utilization of user’s contexts is conducted in generating music recommendation. Yet, developments in music recommender systems lack of utilization of multiple kinds of contexts. Hence, this thesis about development of multi-context-aware music recommender system is conducted. This research elaborate about how to create music recommender systems that is able to utilize more than one user’s context. User’s contexts which are being used here are location context – which can be fetched explicitly from user – and weather by implicitly getting it from the user. Recommendation calculation is done by using Bayesian network with addition of latent variables which are generated based on listening history. Recommender system also utilizes song feature extraction and generate audio frames/words which are also used for recommendation score calculation. Based on the research conducted, it is known that multi-context-aware music recommender system developed is able to give quite a good recommendation. Likelihood of the top 10 songs being picked is quite good with a mean of 0.05. The recommender system is also able to fulfill almost all of the goal of recommender system, except “increasing diversity” as the recommendation generated isn’t very diverse. Besides of that, the recommender system developed is also able to give music recommendation that fits its rank and matches users’ preferences.