Perceptual relativity-based local hyperplane classification

The k-local hyperplane distance nearest neighbors classification (HKNN) builds a non-linear decision surface with maximal local margin in the input space, with invariance inferred from the local neighborhood rather than the prior knowledge, so that it performs very well in many applications. However, it still cannot be comparable with human being in classification on the noisy, the sparse, and the imbalance data. This paper proposes a new approach,called relative local hyperplane classifier(RLHC),to overcome this problem by utilizing the perceptual relativity to HKNN. It finds k nearest neighbors for the query sample from each class and then performs the relative transformation over all these nearest neighbors to build the relative space. Subsequently, each local hyperplane is constructed in the relative space, which is then applied to perform the classification. Experimental results on both real and simulated data suggest that the proposed approach often gives the better results in classification and robustness.