Representation and feature selection using multiple kernel learning

A Novel Multiple Kernel Learning Framework for Heterogeneous Feature Fusion and Variable Selection

“…2, it determines the weight coefficients for each for improved performance while those weights indicate the relevance of the associated features for the learning task. Although the weighted sum of these kernels calculated from individual variable is expected to improve the classification performance, results reported in previous works such as [13] did not achieve significant improvements on several benchmark datasets. Moreover, existing variable selection methods usually regard all variables from the same domain, and the distributions of each variable are assumed to be the same with some data normalization techniques applied.…”

“…A more flexible learning model using multiple kernels instead of one, which is known as multiple kernel learning (MKL), has recently been proposed [12]. Since MKL is able to better represent or discriminate between data using multiple base kernels, it has been shown to improve the performance of many learning tasks, including feature fusion (e.g., [3], [6], [9]) and variable selection (e.g., [13], [14]). …”

“…It aims at identifying a subset of relevant features for improved or comparable recognition performance. Prior works such as [13], [14] have applied MKL for variable selection problems. Dileep et al [13] proposed to learn the optimal base kernels, while each is built from each feature attribute/dimension.…”

“…Prior works such as [13], [14] have applied MKL for variable selection problems. Dileep et al [13] proposed to learn the optimal base kernels, while each is built from each feature attribute/dimension. As a result, this can be regarded as an extreme case of feature fusion.…”

“…MKL also has recently been applied for variable selection [3], [13], [14]. Existing methods typically approach this type of problem as solving a task of learning the optimal weights for each feature variable/attribute.…”