Semi-Supervised Content Based Music Recommendation

Nowadays, many online music stores operate on the World-Wide-Web. The music collection of such stores is approaching the scale of ten million tracks and this has posed a major challenge for searching, retrieving, and organizing music content. Many studies are done in content-based music information retrieval to overcome such difficulty. A promising approach which recently gained much attention is to train a classifier system to predict user priorities and recommend the appropriate music based on this prediction. Although various types of classifiers are incorporated into recommender systems, however, a classifier needs many labelled songs to reach an admissible recommendation. Thus, user should listen to countless songs and provide his opinion for each of them beforethe recommender system could effectively propose any appropriate song to him. In this paper, we offer a Semi Supervised Music Recommender (SS-MR) system which is able to recommend suitable songs with few number of labelled samples. In addition, the system can gradually adapt itself with new songs being added to the system or the user priorities changes over time. Thus training of system from scratch is not needed. Finally extensive experiments are performed with UCI and real world music datasets to evaluate the effectiveness of SS-MR system