Development of an Optimized Tabular Neural Network–Based Predictive Model for a Customer Information Profiling System

Abstract

The rapid expansion and complexity of customer data in modern business and financial environments demand more efficient and accurate profiling techniques. Existing loyalty prediction systems often fail to capture intricate interdependencies within datasets, resulting in suboptimal segmentation and reduced predictive accuracy. Traditional approaches—such as logistic regression, decision trees, clustering algorithms, and standard Tabular Neural Networks (TNN)—also face high computational costs and prolonged training times when handling large-scale data. This study proposes an optimized TNN model integrated with a Modified Augmented Attention Mechanism (MAAM) to overcome these limitations. The MAAM employs a dynamic feature-weighting architecture that enhances efficiency and mitigates latency issues inherent in the original Augmented Attention Mechanism (AAM). The developed MAAM-TNN was evaluated on 4,990 customer records containing 37 demographic and socioeconomic attributes from selected commercial banks. Empirical Results show that MAAM-TNN outperforms both AAM-TNN and standard TNN models, achieving 96.48% specificity, 98.80% sensitivity, 98.30% accuracy, 99.12% precision, a 3.52% false positive rate, and a computational time of 5.99 seconds. The findings demonstrate that MAAM-TNN significantly enhances predictive accuracy and computational efficiency in customer profiling and intelligent financial analytics.