A Study of Personalized Music Recommendation System Using Deep Learning Approaches

With the growing use of internet in recent years, users prefer to listen to music on mobile or tablet via music streaming services. The music streaming services applications provide users with flexibility of listening or downloading the music from the internet. These music streaming service providers (e.g., Spotify, iTunes, KKBOX, etc) usually have a huge amount of music online and make their clients hard to find their favorite songs based on their preferences from corresponding database. Due to this, a music recommendation algorithm is needed according to users’ preference. Digital music data consist of audio signals and textual information describing the music itself. The traditional music recommendation algorithms use the textual information of music for providing recommendations. These traditional approaches suffer with the cold start and data sparsity problems, where they fail to recommend new songs or unpopular songs to the users. We found that user preferences change with respect to the time point of a day, his location, and his social environment. Hence, in this project, we intend to investigate a machine learning based music recommender system which has the capability to analyze users’ preferences with respect to different types of music. In this project, we propose a latent factor model for recommendation named as Personalized Music Recommendation System (PMRS). The PMRS distributes the music data in a cloud environment based on the log files (which contains the transactional details of users and songs), audio signals of the music, and the social data of the users. The social data is collected by crawling the user’s social media account. As the data is stored in three different formats, the PMRS uses three different types of machine learning algorithms: Weighted Feature Extraction (WFE), Gated Recurrent Unit based Recurrent Neural Networks (GRU-RNN) and Long Short Term Memory (LSTM). We will use WFE technique, for generating latent factor vectors for user-songs relationships from the log files. Deep learning based GRU-RNN is used for generating the latent factor vectors from the audio signals. Deep Learning based LSTM technique will be used for generating latent factor vectors from the social data available for the user. Once the latent factor vectors from all the three techniques are generated, we use ensemble technique called Stacking for combining all the latent factor vectors. Finally, based on these combined latent factor vectors, we can recommend music matching users’ preferences. We evaluate the PMRS with existing machine learning techniques with the Million Song Dataset. Preliminary results show that GRU-RNN for generating latent factor vectors produces better performance when we increase the number of hidden layers in the deep neural network architecture.

Project ID:PB10703-1493
External Project ID:MOST106-2221-E182-065