Playlist-Based Tag Propagation for Improving Music Auto-Tagging

Abstract

The performance of a music auto-tagging system highly relies on the quality of the training dataset. In particular, each training song should have sufficient relevant tags. Tag propagation is a technique that creates additional tags for a song by passing the tags from other similar songs. In this paper, we present a novel tag propagation approach that exploits the song coherence of a playlist to improve the training of an auto-tagging model. The main idea is to share the tags between neighboring songs in a playlist and to optimize the auto-tagging model through a multi-task objective function. We test the proposed playlist-based approach on a convolutional neural network for music auto-tagging and show that it can indeed provide a significant performance improvement.