Similarity Learning via Optimizing the Data-Dependent Kernel

In this paper, we present a scheme of similarity measure learning based on kernel optimization. Employing a data-dependent kernel model, the proposed scheme optimizes the spatial distribution of the training data in the feature space, aiming to maximize the class separability of the data in the feature space. The learned similarity measure, derived from the optimized kernel, exhibits a favorable feature to the task of pattern classification, that the spatial resolution of the embedding space is expanded around the boundary areas, and shrunk around the homogeneous areas. Experiments demonstrate that using the learned similarity measure can substantially improve the performances of the K-nearest-neighbor classifier.