Information Fusion VIA Optimized KECA with Application to Audio Emotion Recognition

As a recent proposed information fusion tool, Kernel Entropy Component Analysis (KECA) has attracted more attentions from the research communities of multimedia. It utilizes descriptor of entropy estimation and achieves improved performance for information fusion. However, KECA roughly reduces to sorting the importance of kernel eigenvectors by entropy instead of by variance as in Kernel Principal Components Analysis (KPCA), without extracting the optimal features retaining more entropy of the input data. In this paper, a novel approach Optimized Kernel Entropy Components Analysis (OKECA) is introduced to information fusion, which can be considered as an alternative method to KECA for information fusion. Since OKECA explicitly extracts the optimal features that retain most informative content, it leads to improving the final performance or classification accuracy. To demonstrate the effectiveness of the proposed solution, experiments are conducted on Ryerson Multimedia Lab (RML) and eNTERFACE emotion datasets. Experimental results show that the proposed solution outperforms the existing methods based on the similar principles, and the Deep Learning (DL) based method.