Unveiling Customer Insights: An Interpretable Machine Learning Approach to Bank Telemarketing Data

Abstract

Financial institutions play a vital role in driving the economy. Despite the advent of digital financial systems, phone-based product offerings remain popular in the banking sector. This study focuses on building a model to predict customer applications from telemarketing campaigns. By utilizing the publicly available Bank Marketing Data Set, which exhibits significant class imbalance, we explored various combinations of effective imbalanced treatments and categorical encodings in conjunction with machine learning models to identify the most optimal combination for prediction. Additionally, interpretable machine learning techniques were employed to delve into the critical features and the underlying reasoning behind the model's predictions. The experiment revealed that the LightGBM model with Class weight and One-hot encoding yielded the best AUC score of 0.948. Using SHAP to explain the model's behavior, we found that the features related to economic factors hold greater significance compared to individual customer attributes. Furthermore, error analysis on false negative instances demonstrated that the similarity of instance characteristics of some important features could mislead the models and result in inaccurate prediction. These findings shed light on the model's decision-making process and offer insights for enhancing prediction accuracy and understanding customer behavior in financial product applications. The results offer actionable guidance for optimizing business operations by enabling more efficient lead targeting, reducing resource waste in telemarketing efforts, and supporting data-driven decision-making in customer outreach strategies.