Individual Identification using Third-Order Tensor-Based Multilinear FisherECG

Abstract

In this study, we present individual identification using a third-order tensor-based multilinear FisherECG (MFECG). The MFECG is based on multilinear linear discriminant analysis (LDA) in the field of multilinear subspace learning (MSL). MSL performs feature extraction without rearranging input data into 1-D. Thus, the MSL can maintain the original correlation in structure. Also, since the compact features are extracted, the large tensor can be efficiently handled. However, linear subspace learning (LSL) must reorder the data into 1-D for feature extraction. That makes the dimension of the data become very large, which causes the problem of insufficient number of training data. Here, the 3-D tensor is formed by rearranging a 1-D electrocardiogram signal into a 2-D matrix and accumulating the 2-D matrices. The experimental results show that the tensor-based MFECG showed higher accuracies than PCA (Principal Component Analysis) and LDA of LSL.