Joint diagonalization of kernels for information fusion

Abstract

Information Fusion is becoming increasingly relevant in fields such as Image Processing or Information Retrieval. In this work we propose a new technique for information fusion when the sources of information are given by a set of kernel matrices. The algorithm is based on the joint diagonalization of matrices and it produces a new data representation in an Euclidean space. In addition, the proposed method is able to eliminate redundant information among the input kernels and it is robust against the presence of noisy variables and irrelevant kernels. The performance of the algorithm is illustrated on data reconstruction and classifications problems. © Springer-Verlag Berlin Heidelberg 2007.