Audio-Based Sequential Music Recommendation

Abstract

We propose an audio-based recommendation model designed to predict the upcoming track within a listening session, given the audio associated with the current track. Instead of relying on users' feedback, as most recommenders, the proposed model aims to learn intrinsic audio elements that can be leveraged in the context of sequential recommendation. The proposed model is evaluated using Mel-spectrogram and raw audio as input data and, in its best configuration, was able to predict almost 65% unseen transitions used in the evaluation phase, and 3.5% cold-start transitions, i.e. transitions from tracks that were never seen by the model.