| We apply a new machine learning tool, kernel combination, to the task of semantic music retrieval. We use 4 different types of acoustic content and social context feature sets to describe a large music corpus and derive 4 individual kernel matrices from these feature sets. Each kernel is used to train a support vector machine (SVM) classifier for each semantic tag (e.g., "aggressive", "classic rock", "distorted electric guitar") in a large tag vocabulary. We examine the individual performance of each feature kernel and then show how to learn an optimal linear combination of these kernels using semi-definite programming. We find that the retrieval performance of the SVMs trained using the combined kernel is superior to SVMs trained using the best individual kernel for a large number of tags. In addition, the weights placed on individual kernels in the linear combination reflect the relative importance of each feature set when predicting a tag. |
| The graphic below illustrates the distribution of weight given to the 4 feature kernels in the optimum combined-kernel SVM for each of the 61 words used: |
|
|
Subset of 61 words from the CAL500 vocabulary used for kernel combination experiments
List of synonyms used to map CAL500 tags to synonymous last.fm tags
Relevant Publications
| L. Barrington, D. Turnbull, M. Yazdani, G. Lanckriet - Combining Audio Content and Social Context for Semantic Music Discovery. SIGIR, 2009. |
| Barrington, Yazdani, Turnbull and Lanckriet. Combining Feature Kernels for Semantic Music Retrieval. ISMIR 2008. |
| Lanckriet, Cristianini, Bartlett, El Ghaoui, and Jordan. Learning the kernel matrix with semi-definite programming. Journal of Machine Learning Research, 5:27-72, 2004 |
| You can download Guillaume Obozinski's MATLAB code to implement the kernel combination algorithm. |