Adaptive One-Dimensional Convolutional Neural Network for Tabular Data

Abstract

This study introduces an innovative approach for tackling the credit risk prediction problem using an Adaptive One-Dimensional Convolutional Neural Network (1D CNN). The proposed methodology is designed for one-dimensional data, such as tabular data, through a combination of feed-forward and back-propagation phases. During the feed-forward phase, neuron outputs are computed by applying convolution operations to previous layer outputs, along with bias terms and activation functions. The subsequent back-propagation phase updates weights and biases to minimize prediction errors. A custom weight initialization algorithm tailored to Leaky ReLU activation is employed to enhance model adaptability. The core of the proposed algorithm lies in its ability to process each training data sample across layers, optimizing weights and biases to achieve accurate predictions. Comprehensive evaluations are conducted on various machine learning algorithms, including Gaussian Naive Bayes, Logistic Regression, ensemble methods, and neural networks. The proposed Adaptive 1D CNN emerges as the top performer, consistently surpassing other methods in precision, recall, F1-score, and accuracy. This success is attributed to its specialized weight initialization, effective back-propagation, and integration of 1D convolutional layers.