Music recommendation system

Abstract

Traditional music recommendation techniques are based on the concept of collaborative filtering (CF) which leverages the listening patterns of many users. While this technique is very effective, it is not able to recommend new songs or artists since there is no listening history to draw upon. This research addresses the issue of music recommendation for novel songs and artists by making recommendations which are solely based on the audio content rather than metadata such as genre, artist, or using the listening histories of other users. We propose a similarity metric based solely on raw audio content, which is then used to recommend songs that are close to the user’s library, based on this similarity function. The similarity metric was created by training a Siamese Neural Network (SNN) on a dataset of similar and dissimilar song pairs. Each song was first converted into a Mel-Spectrogram to obtain its bitmap representation. The SNN consists of two identical CNNs which are fed the Mel-Spectrogram bitmap of each song pair. By training the model on this dataset of song image pairs, the SNN learns to act as a similarity metric between songs based, on the raw audio content. The SNN model achieved an accuracy score of 81.64% on the test set. A query-by-multiple-examples based music recommendation system was created, which makes use of the developed similarity metric to make music recommendations. In order to evaluate the performance of our proposed system, a survey website was developed. Survey participants first create a preference set containing songs they enjoy, after which they proceed to rate recommendations. A naive genre-based baseline system was also implemented, and the recommendations made by both systems are interleaved randomly, after which the user rates a single list of songs. The survey is a blind-use study, thus participants are not aware of the fact that two systems are being used, which helps to eliminate possible biases. The findings showed that the majority of the participants preferred the recommendations made by the proposed system, which also received significantly higher ratings than the baseline system, whilst recommending songs that are less popular than those recommended by the baseline system.