Background Structural analysis in business and econometric systems traditionally relied on numerical proportionality-based modeling techniques. However, the increasing dimensionality and heterogeneity of modern business datasets have exposed limitations in conventional dimensionality reduction approaches such as Principal Component Analysis (PCA), particularly in preserving relationships between input variables and target outcomes. Deep learning approaches such as Convolutional Neural Networks (CNNs) have recently emerged as alternatives capable of automated feature extraction and adaptive structural learning. Methods This study evaluated the effectiveness of a one-dimensional CNN (1D-CNN) for structural analysis of business data, using customer churn prediction as a case study, and comparatively assessed its performance against PCA-supported Artificial Neural Networks (PCA-ANNs). A customer churn dataset containing 440,882 customer records was obtained from Kaggle. The dataset consisted of numerical and categorical business variables associated with customer attrition. A preprocessing pipeline involving custom StringLookup implementations, one-hot encoding, categorical indexing, standardization, and PCA transformations was developed for the respective models. The dataset was divided into training, validation, and testing sets using a 64:16:20 ratio. A custom weighted binary cross-entropy loss function was introduced to address class imbalance during training. Multiple PCA-ANN models corresponding to varying principal component sizes were evaluated alongside the proposed 1D-CNN architecture. Model evaluation was conducted using accuracy, precision, recall, F1-score, and area under the curve. Results The proposed 1D-CNN model achieved a training accuracy of 98.50%, a validation accuracy of 98.29%, and a test accuracy of 98.34%, outperforming PCA-ANN models utilizing up to eight principal components. The highest overall performance was achieved by the PCA-ANN-9 model, which attained a training accuracy of 99.25%, a validation accuracy of 99.12%, and a test accuracy of 99.13%. Despite its slightly lower predictive accuracy, the CNN architecture demonstrated advantages in automated supervised feature extraction, adaptability to heterogeneous business data, and reduced dependence on manually selected dimensionality reduction parameters. Conclusion The findings demonstrate that 1D-CNN architectures provide an effective and scalable alternative for the structural analysis of business data. While the final PCA-supported ANN model achieved marginally higher predictive accuracy in this study, the CNN model offered greater flexibility, automated feature learning, and enhanced adaptability to evolving business environments. The study further highlights the potential of deep learning-based structural modeling approaches for customer behavior analysis and predictive business analytics.
Daniel A Folorunso (Tue,) studied this question.