融合深度特征与多核学习的LSTWSVM及其工业应用

为了提高多核学习(MKL)的表示能力同时降低其计算成本,提出一种融合深度特征与多核学习的最小二乘孪生支持向量机(LSTWSVM)算法.针对支持向量机等核分类器在多核学习中高计算复杂度的问题,提出一种基于边缘错误最小化原则的多核LSTWSVM框架,利用分类器优势提高多核学习的性能.针对高斯多核浅层结构的问题,采用MKL法设计一种基于深度神经网络多层信息的高鲁棒性深度映射核,将此深度核与多尺度高斯基核以核矩阵哈达玛积方式相融合,构造一组新的具有高度表达能力的改进核.最后,将基于LSTWSVM的多核训练算法与改进的多核结构进行高度集成,通过大量基准数据集与工业数据实验表明,其能有效结合深度学习与多核学习的优势,且以较低的计算成本提高分类精度与泛化能力.

LSTWSVM fusion of deep feature and multiple kernel learning and its industrial applications

In order to improve the representation ability of multi-kernel learning and reduce its computational cost, this paper proposes a least squares twin support vector machine(LSTWSVM) algorithm that combines depth feature and multi-kernel learning. Aiming at the problem of high computational complexity of kernel classifiers such as support vector machine in multi-kernel learning, a multi-kernel LSTWSVM framework based on the principle of edge error minimization is proposed. The cost-sensitive learning idea is adopted to improve the performance of multi-kernel learning by using the advantages of classifiers. Aiming at the problem of Gauss multi-kernel shallow structure, a highly robust depth mapping kernel based on depth neural network multi-layer information is designed using the MKL method. The depth kernel and multi-scale Gaussian basis kernel are fused in the form of kernel matrix Hadamard product to construct a new set of improved cores with high expressiveness, which contains the deep feature information of data. Finally, this paper highly integrates the multi-kernel training algorithm based on the LSTWSVM with the improved multi-kernel structure. Through benchmark datasets and industrial experiments, it shows that it can combine the advantages of deep learning and multi-kernel learning, and improve the classification accuracy and generalization ability at a lower computational cost.