Designing an efficient gradient descent based heuristic for clusterwise linear regression for large datasets

Abstract

Multiple linear regression is the method of quantifying the effects of a set of independent variables on a dependent variable. In clusterwise linear regression problems, the data points with similar regression estimates are grouped into the same cluster either due to a business need or to increase the statistical significance of the resulting regression estimates. In this paper, we consider an extension of this problem where data points belonging to the same category should belong to the same partition. For large datasets, finding the exact solution is not possible and many heuristics requires an exponentially increasing amount of time in the number of categories. We propose variants of gradient descent based heuristic to provide high-quality solutions within a reasonable time. The performances of our heuristics are evaluated across 1014 simulated datasets. We find that the comparative performance of the base gradient descent based heuristic is quite good with an average percentage gap of 0.17 % when the number of categories is less than 60. However, starting with a fixed initial partition and restricting cluster assignment changes to be one-directional speed up heuristic dramatically with a moderate decrease in solution quality, especially for datasets with a multiple number of predictors and a large number of datasets. For example, one could generate solutions with an average percentage gap of 2.81 % in one-tenth of the time for datasets with 400 categories.