Feature extraction of multimodal data by cluster-based correlation discriminative analysis

Linear discriminant analysis (LDA) is suboptimal in dealing with multimodal data that multiple clusters per class exist in input space. This is caused by its inherent globality. To attack this problem, a novel extension of LDA is presented which is called cluster-based correlation discriminative analysis (CCDA). CCDA encodes correlation-based similarity metric in cluster structure modeling, aiming to preserve the correlational affinity in lower-dimensional subpace. Extensive experiments on two widely used databases validate that CCDA outperforms existing LDA variants in facial expression recognition tasks.