Binary Classification from Interval Constraint Learning

Abstract

In this article, we present a novel method for binary classification based on interval constraint solving. A binary classification problem in our framework is formulated as a constraint-solving problem, where constraints are linear equations and membership to intervals. The framework adopts fuzzy logic with two goals. The first is that the datasets are mapped to fuzzy sets in such a way that the data coordinates are rescaled to the interval [0,1]. Additionally, fuzzy set mapping enables redistributing data in the fuzzy space. A constraint solver CLP(BNR) is used to solve the constraints, which acts as a learning method. A classifier in our framework is a solution to the constraints problem consisting of a set of interval constraints expressing the bounds for the coefficients of linear equations and a set of fuzzy sets. The underlined classification procedure requires checking the satisfiability of the interval constraints from the mapping of data into fuzzy sets. An ensemble of the learning method is also proposed to improve classification.