In this work we present a novel approach for supervised multiclass classification of temporal signals. We present two core contributions. First, we derive a local data embedding we denote as Local Diffusion-PCA, which encodes the local latent low-dimensional manifold structure of temporal signal segments. Second, we propose to utilize the similarity in classes that is characteristic of common signal classification problems, and derive a classifiers sharing approaches that utilize this property. The first of which is based on AdaBoost and the second on multilayered stacked SVM. The proposed scheme is applied to musical genres classification, as this problem was thoroughly studied and thus provides an algorithmic testbed, consisting of contemporary datasets and results to compare against. Our approach is shown to compare favorably with state-of-the-art approaches.
Automatic Large Margin Music Genre Classification in Hierarchical Local Diffusion Spaces*
Yosi Keller - Bar-Ilan University, Israel
Sep 11 2014 - 3:00pm
110 Fine Hall