Matrix-pattern-oriented least squares support vector classifier with AdaBoost

Abstract

Matrix-pattern-oriented least squares support vector classifier (MatLSSVC) can directly classify matrix patterns and has a superior classification performance than its vector version least squares support vector classifier (LSSVC) especially for images. However, it can be found that the classification performance of MatLSSVC is matrixization-dependent, i.e. heavily relying on the reshaping ways from the original (vector or matrix) pattern to (another) matrix. Thus, it is difficult to determine which reshaping way is fittest to classification. On the other hand, the changeable and different reshaping ways can naturally give birth to a set of MatLSSVCs with diversity and it is the diversity that provides a means to build an ensemble of classifiers. In this paper, we exactly exploit the diversity of the changeable reshaping ways and borrow AdaBoost to construct an AdaBoost–MatLSSVC ensemble named AdaMatLSSVC. Our contributions are that: (1) the proposed AdaMatLSSVC can greatly avoid the matrixization-dependent problem on single MatLSSVC; (2) different from the ensemble principle of the original AdaBoost that uses a single type of classifiers as its base components, the proposed AdaMatLSSVC is on top of multiple types of MatLSSVCs in different reshapings; (3) since AdaMatLSSVC adopts multiple matrix representations of the same pattern, it can provide a complementarity among different (matrix) representation spaces; (4) AdaMatLSSVC mitigates the selection of the regularization parameter, which are all validated in the experiments here.