Fast Class-Attribute Interdependence Maximization (CAIM) Discretization Algorithm

Discretization is a process of converting a continuous attribute into an attribute that contains small number of distinct values. One of the major reasons for discretizing an attribute is that some of the machine learning algorithms perform poorly with continuous attribute and thus require front-end discretization of the input data. The paper describes a Fast Class-Attribute Interdependence Maximization (F-CAIM) algorithm that is an extension of the original CAIM algorithm. The algorithm works with supervised data by maximization of the classattribute interdependence. The F-CAIM’s improvement of the CAIM algorithm is significant shortening of the computational time required to discretize the data. It has all CAIM’s advantages like fully automated generation of possibly minimal number of discrete intervals, achieving the highest class-attribute interdependency when compared with other discretization algorithms, and improving performance of machine learning algorithms that are subsequently used on the discretized data. We present the results based on extensive benchmarking tests of F-CAIM, CAIM and six other state-of-the-art discretization algorithms. The tests use eight wellknown machine learning datasets consisting of continuous and mixed-mode attributes. They show that the F-CAIM’s speed is comparable to the speed of the simplest unsupervised algorithms and better than these of other supervised discretization algorithms.