Local Feature Selection by Formal Concept Analysis for Multi-class Classification

Abstract

In this paper, we propose a multi-class classification algorithm to apply it to data sets increasing frequently. The algorithm performs lazy learning based on formal concept analysis. We designed it so that it obtains localness in predicting classes of test data and feature selection simultaneously. From a given data set that consists of a set of training data and a set of test data, the algorithm generates a single formal concept lattice. Every formal concept in the lattice represents a cluster of data that are generated by various feature selections. In order to classify each test datum, plausible clusters are selected and combined into a set of neighbors for the test datum. Our algorithm can construct sets of neighbors for test data that are never generated by other algorithms, e.g., the $k$-nearest neighbor algorithm and decision tree classifiers. We compare our algorithm with other algorithms by experiments using UCI datasets and show that ours is comparable to the others at the viewpoint of correctness.