Multi-model telecom churn prediction

Abstract

Customer turnover is likely to be significant in the telecom business due to its dynamic and competitive nature. Traditional measures of performance are inadequate in such a fluid environment to accurately portray organisational objectives. The reason behind this is because the performance measurements are not in line with the company goals. A multi-model telecom churn prediction (MMTCP) with minority upliftment techniques is presented in this work. It can handle data imbalance successfully and has a loss function that separates loss into two parts: loss due to incorrect prediction and loss due to unavoidable loss. First, utilising a set of training data and a number of diverse base learners, MMTCP generates predictions at the first level; second, using these predictions as a starting point, it uplifts the minority in the model. Gradient-boosted trees and naïve bayes make up the first stage, while one-class SVM is the basis of the second combiner stage. As compared to both current classifier models and the state-of-the-art churn prediction methods from literature, the experimental findings suggest that the MMTCP model exhibits 1 to 7% greater churn prediction levels and 1.3 to 1.7 times decreased loss levels.