*Canh Hao Nguyen, Tu Bao Ho*

We study the problem of evaluating the goodness of a kernel matrix for a classification task. As kernel matrix evaluation is usually used in other expensive procedures like feature and model selections, the goodness measure must be calculated efficiently. Most previous approaches are not efficient, except for Kernel Target Alignment that can be calculated in O(n2) time complexity. Although KTA is widely used, we show that it has some serious drawbacks. We propose an efficient surrogate measure to evaluate the goodness of a kernel matrix based on the data distributions of classes in the feature space. The measure not only overcomes the limitations of KTA, but also possesses other properties like invariance, efficiency and error bound guarantee. Comparative experiments show that the measure is a good indication of the goodness of a kernel matrix.