Improved Neural Network Predictions with Correlation-Based Subset Selection

This paper introduces a method called Artificial Neural Network with Best Subset Selection (ANNBS) to improve the accuracy of predictions in neural networks. The method involves selecting an optimum subset of variables. Although neural networks have great prediction capabilities, they do not possess inherent processes for selecting variables. This approach uses neural networks to choose variables, employing a Monte Carlo algorithm to conduct subset searching across 1,000 iterations. The analysis encompasses both actual and simulated datasets obtained from neural network models, with correlation coefficients ranging from 0 to 0.8. The correlation coefficient is used as the primary criterion for selecting the optimal subgroup. This strategy is compared to linear regression, traditional neural networks, a hybrid model that combines stepwise regression analysis with artificial neural networks, and model averaging strategies that use information criteria and fit-based weights. The findings indicate that the ANNBS approach consistently delivers low root mean square errors (RMSEs) for both the training and test data, even when the sample sizes are small. As the number of samples increases, the disparity between the RMSEs of the training and test data decreases.