Churn Prediction Using Mathematical Programming via a Linear Classification Algorithm

Abstract

The initial idea in developing customer churn prediction was to use statistical analysis of a sample of previous customers to help operators predict which existing customers were likely migrate to other operators. In this paper, a nonstatistical approach to this problem can also be envisaged by formulating the problem as a linear program. The main concept of our proposed algorithm is to hybridize the Jackknife resampling technique with an efficient heuristic based on a variable neighborhood search algorithm to construct a new approach that can present an alternative to black-box machine learning methods. The goal is to find possible solutions (or weights) to minimize error distances and the number of misclassified points. We tested the suggested algorithm with an imbalanced instance and the comparisons show that the proposed model outperforms the most existing machine learning classifiers, both in terms of the solution quality and the execution time. The numerical results indicate that the best solution is achieved in 85% over G-mean of all instances tested. The average gap between our solution and the best solution is still quite small (1.15%).