Perceptual nearest neighbors for classification

Finding nearest neighbors plays a fundamental role in many artificial intelligence tasks, such as manifold learning, data mining, and information retrieval, etc. Directly applying this idea to perform classification is simple and often results in good performance on complex data types. However existing classifiers apply a well designed measure to find nearest neighbors. They still can not be comparable with human being in many complex cases such as on noisy, sparse or high dimensional data. This paper proposes a quite different but much interesting approach that utilizes Lipschitz function to define a simple topological transformation for modeling Gestalt laws of psychology from data and then designs a new measure to evaluate the quality of the discovered Gestalts. Subsequently, the nearest neighbors are selected from higher quality Gestalts, from which a new classifier is proposed that has much better classification performance.