Improved Session-Based Recommendations Using Recurrent Neural Networks for Music Discovery
Yi Ye, Yi Fei Xie, Cheng Chen · Journal of Physics Conference Series · 2019
Abstract The number of songs on the music platform is large while users just have access to a fraction of songs. Recommendation systems are created to solve this problem and collaborative filtering is the most widely used method. But collaborative filtering methods are overly dependent on user ratings of projects. New recommendation methods, Deep Learning techniques, are proposed especially for items’ feature extraction and for session-based recommendation with RNN. In this paper, we considered a LSTM-based approach for session-based recommendation and built architecture for music discovery. This architecture is composed of four modules: MDM (Music Data Modeling), NSP (Next Song Prediction Using LSTM), MLB (Music Library Building) and R4U (Recommendation for Users). Improved session-parallel mini-batches and ranking loss function are leveraged by the architecture to modify the basic LSTM. Through the experiment, we obtained a relative improvement in ranking metrics (MRR@N and Recall@N) over the session-based algorithms baselines.