Discretization Algorithm that Uses Class-Attribute Interdependence Maximization

摘要

Most of the existing machine learning algorithms are able to extract knowledge from databases that store discrete attributes (features). If the attributes are continuous, the algorithms can be integrated with a discretization algorithm that transforms them into discrete attributes. The paper describes an algorithm, called CAIM (class-attribute interdependence maximization), for discretization of continuous attributes that is designed to work with supervised learning algorithms. The algorithm maximizes the class-attribute interdependence and, at the same time, generates possibly minimal number of discrete intervals. Its big advantage is that it does not require the user to pre-define the number of intervals, in contrast to many existing discretization algorithms. The CAIM algorithm and five other state-of-the-art discretization algorithms were tested on well-known machine learning datasets consisting of continuous and mixed-mode attributes. The tests show that the proposed algorithm generates discrete attributes with, almost always, the highest class-attribute interdependency when compared with other algorithms, and at the same time it always generates the lowest number of intervals. The discretized datasets were used in conjunction with the CLIP4 machine learning algorithm. The accuracy of the rules generated by the CLIP4 shows that the proposed algorithm significantly improves classification performance; it also performs best in comparison with other five discretization algorithms. The CAIM algorithm's speed is comparable to the simplest unsupervised algorithms and outperforms other supervised discretization algorithms.