Tag Propagation and Cost-Sensitive Learning for Music Auto-Tagging

Abstract

The performance of music auto-tagging depends on the quality of training data. In practice, the links between songs and tags in the manually labeled training data can be incorrect (false positive) or missing (false negative). In this paper, we propose a cost-sensitive tag propagation learning method to improve auto-tagging. Specifically, we exploit music context to determine similar songs and propagate tags between them. Both propagated tags and original tags are used to optimize the auto-tagging models, and cost-sensitivity is incorporated into the loss function to enhance the robustness by adjusting the weight of relevant ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">positive</i> ) links with respect to irrelevant ( <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">negative</i> ) links. The proposed method is tested on three auto-tagging models: 2D-CNN, CRNN, and SampleCNN. The Million Song Dataset is used for training, and four music contexts, artist, playlist, tag, and listener, are used for song similarity measurement. The experimental results show 1) The proposed method can successfully improve the performance of the three auto-tagging models, 2) The cost-sensitive loss function helps reduce the impact of missing tags, and 3) The artist music context is more powerful for tag propagation than the other three music contexts.