Music Recommendation System based on Collaborative Filtering and Singular Value Decomposition

Abstract:
The most significant thing for the recommendation system is how to understand the user's portrait. The idea is to integrate the data to learn exactly what music users like. Nowadays, music recommendation systems give chances to related industries to attract and aggregate users. Many recent models of music recommendation are proposed. The algorithms used include flow log, sorting net, and linear regression. They are not efficient and accurate enough when the information of music data is single-featured. A fairly complete music recommendation system is built with several calculable algorithms in this paper. The algorithm Collaborative Filtering is based on users (User CF) and items (Item CF) respectively and Singular Value Decomposition (SVD) is used in the music recommendation system. While using the Item CF, calculating the similarity between user songs and all unique songs in the training data to generate the Jaccard index to make a comparison. The User CF is to calculate the similarity between certain users playing certain songs and all users in the training data to generate the Jaccard index as a standard. The SVD algorithm decomposes the feature matrixes of the rate of certain users and songs. By comparing the ACC, TPR, FNR, and other indicators which are 92%, 94.12%, and 5.88%, the SVD is more accurate and efficient than the other two algorithms. The User CF and Item CF also maintain high accuracy in the prediction which is 88% and 80%. The work demonstrates that when there is a single feature, SVD and CF algorithms can help make predictions through available data.