Improving Musical Concept Detection by Ordinal Regression and Context Fusion

To facilitate information retrieval of large-scale music databases, the detection of musical concepts, or auto-tagging, has been an active research topic. This paper concerns the use of concept correlations to improve musical concept detection. We propose to formulate concept detection as an ordinal regression problem to explicitly take advantage of the ordinal relationship between concepts and avoid the data imbalance problem of conventional multi-label classification methods. To further improve the detection accuracy, we propose to leverage the co-occurrence patterns of concepts for context fusion and employ concept selection to remove irrelevant or noisy concepts. Evaluation on the cal500 dataset shows that we are able to improve the detection accuracy of 174 concepts from 0.2513 to 0.2924.