Class-wise feature extraction technique for multimodal data

Within-class multimodality happens when the scattering of the patterns having the same class label follows more than one modal distribution. In this multimodal scenario, it is important to preserve the intrinsic information of the classes when reducing the dimensionality of the data. However, many feature extraction techniques are incapable of dealing properly with this kind of scenario. This paper proposes a method called Class-dependent Locality Preserving Projections that operates in scenarios that present within-class multimodality. Class-dependent Locality Preserving Projections evaluates each class separately, creates a specific projection for each one of them, analyzes a query pattern using the output of each class and classifies based on the class that better fits the pattern. The experimental study involves real and artificial datasets that were created to show within-class multimodality. The results indicate that Class-dependent Locality Preserving Projections can be used as a feature extraction technique for general purposes, and it is particularly advantageous when applied to within-class multimodal scenarios.