MLP Neural Networks vs. Logistic Regression: A Comparative Study of Customer Churn Prediction in Bank Marketing

Abstract

Customer churn is a recurring problem in bank marketing, and there are several machine learning approaches that can help. This paper compares two of them: a Multilayer Perceptron (MLP) neural network and logistic regression. We used the UCI Bank Marketing dataset, which has 41,188 client records from direct marketing campaigns run by a Portuguese bank between 2008 and 2013. Training and evaluation were kept the same for both models. We adjusted class weights to deal with the imbalance, since only about 11 percent of clients subscribed. Performance was measured with precision, recall, F1-score, and ROC-AUC. The MLP did better on the main metrics, specifically recall for the positive class and ROC-AUC, but not by much. Logistic regression actually performed better on overall accuracy and a few other measures. It also has the benefit of being easier to interpret. What this suggests is that on structured tabular data, a more complex model does not automatically translate into better results.