关键词: 核方法 组合核矩阵 特征距离 模型选择

Abstract:

Combinatorial construction is an effective and promising kernel construction approach. We study a criterion for selecting an optimal combinatorial ker   nel matrix in the feature space. Different from the previously popular kernel target alignment criteria, the new criterion overcomes the limitation and  measures the performance of a combinatorial kernel matrix based on feature distance. We first introduce the kernel target alignment and discuss its limi  tation for the combinatorial kernel matrix. Then we present the feature-distance-based combinatorial kernel matrix evaluation criterion formally. Finall   y, we analyze the properties of the proposed criterion and examine its performance based on the simulated data. Both the theoretical analysis and the ex perimental results demonstrate that the proposed combinatorial kernel matrix evaluation criterion is sound and effective.

Key words: kernel methods;combinatorial kernel matrix, feature distance, model selection