Better Aware of Anticipating Customer Attrition in Telecommunications via Enhanced Tabular Learning Networks

Abstract

In telecom, retaining customers is vital given high acquisition costs. Our study employs TabNet, using sequential attention in tabular data. We introduce alpha parameters, enabling and adjusting sparsity and enhancing model adaptability in transductive training in TabNet's transformer block via α-entmax for improved churn prediction. Addressing data imbalance, we apply SMOTE and random oversampling, plus combined methods like SMOTE-TOMEK, SMOTE-ENN, and SMOTE-RENN. Tested on public telecommunication data, our goal is an accurate churn prediction tool to aid telecoms in retention and cost reduction.