Identification of Linear Regression Models by a Clustering Algorithm

This paper is a contribution to the problem of representation of clusters by a set of regression models.A new divisive clustering algorithm, reffered to as CLUREG, is presented. The algorithm belongs to the class of exchange methods, is able to deal with large sets of units and is such that each cluster of units is “well represented” by a specific regression model. The quality of the representation is measured by the sum of squares of the errors of the fitted regression model.CLUREG, by the k- median algorithm, optimally clusters the variables. Then, after an initial partition of the units, based on the clusters of the variables, reallocates the units by the optimization of the overall sum of squares of errors in the clusters.The computational complexity is a linear function of the number of units and a non linear function of the number of variables.