A Discrete Approach for Supervised Pattern Recognition

摘要

We present an approach for supervised pattern recognition based on combinatorial analysis of optimum paths from key samples (prototypes), which creates a discrete optimal partition of the feature space such that any unknown sample can be classified according to this partition. A training set is interpreted as a complete graph with at least one prototype in each class. They compete among themselves and each prototype defines an optimum-path tree, whose nodes are the samples more strongly connected to it than to any other. The result is an optimum-path forest in the training set. A test sample is assigned to the class of the prototype which offers it the optimum path in the forest. The classifier is designed to achieve zero classification errors in the training set, without over-fitting, and to learn from its errors. A comparison with several datasets shows the advantages of the method in accuracy and efficiency with respect to support vector machines.