A multi-input with multi-function activated weights and structure determination neuronet for classification problems and applications in firm fraud and loan approval

Neuronets trained by a weights-and-structure-determination (WASD) algorithm are known to resolve the shortcomings of traditional back-propagation neuronets such as slow training speed and local minimum. A multi-input multi-function activated WASD neuronet (MMA-WASDN) model is introduced in this paper, combined with a novel multi-function activated WASD (MA-WASD) algorithm, for handling binary classification problems. Using multiple power activation functions, the MA-WASD algorithm finds the optimal weights and structure of the MMA-WASDN and uses cross-validation to address bias and prevent being stuck in local optima during the training process. As a result, neuronets trained with the MA-WASD algorithm have higher precision and accuracy than neuronets trained with traditional WASD algorithms. Applications on firm fraud and loan approval classification validate our MMA-WASDN model in order to demonstrate its outstanding learning and predicting performance. Since these applications use real-world datasets that include strings and missing values, an algorithmic method for preparing data is also suggested to make them manageable from the MMA-WASDN. A comparison of the MMA-WASDN model to five other high-performing neuronet models is included, as well as a MATLAB package that is publicly available through GitHub to support and promote the findings of this research. (C) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper introduces a MMA-WASDN model trained by a WASD algorithm, combined with an MA-WASD algorithm for binary classification problems. The model achieves higher precision and accuracy through multiple power activation functions and cross-validation. Applications in fraud and loan approval validate the outstanding performance of the model, and a comparison with other models is provided.