Abstract
This dissertation introduces several problems on music recommendation. To make a high-qualified recommender, we evaluate feature importance for favorite song detection from two perspectives. The experiment results expose that collaborative filtering signal is the most important feature among the analyzed features. A classifier combination method is proposed to leverage several classifiers trained by different data sources to predict music genre. The complemented genre labels are used in a recommendation system for individual users on local device. The recommender takes freshness, time pattern, genre, publish year, and favor into account to make recommendations. The recommender outperforms the baseline on mostly favorite songs. We propose an adapted recommendation method to response user feedbacks and find out local optimizations to improve the recommendation quality. Furthermore, a probability-based method is proposed to make recommendations for implicit user groups by integrating individual opinions on music.