Evaluation of Deep Learning Models for Tabular Data Classification of Deposit Customers

Abstract

The application of deep learning to tabular data still faces various challenges, especially in the banking domain, which is dominated by numerical and categorical features without a clear temporal structure. This study aims to evaluate and compare the performance of several deep learning architectures in the task of classifying time deposit customers. The models used include Multilayer Perceptron, Convolutional Neural Network, and Long Short Term Memory-based Recurrent Neural Network. The dataset used is banking customer data with a binary classification target, namely subscribing or not subscribing to deposits. The research stages included data preprocessing, model architecture development, training process, and performance evaluation using accuracy, precision, recall, and area under the ROC curve metrics. Additional analysis was performed on the training curve to assess the optimization stability and generalization ability of the model. The experimental results showed that all models were able to achieve an accuracy rate above 89 percent. The Convolutional Neural Network model provides the most balanced performance with the highest accuracy and precision as well as stable training dynamics. Multilayer Perceptron shows good class separation capabilities, but tends to overfit, while Recurrent Neural Network does not provide significant performance improvements and shows instability during training. These findings confirm that the suitability of deep learning architecture to the characteristics of tabular data is an important factor in model selection for banking applications.