Threshold accepting trained principal component neural network and feature subset selection: Application to bankruptcy prediction in banks

This paper proposes an application of new principal component neural network (PCNN) architecture to bankruptcy prediction problem in commercial banks. Further, a new feature subset selection (FSS) algorithm is proposed. In this architecture, the hidden layer is completely replaced by what is referred to as a ‘principal component layer’. This layer consists of a few selected principal components that perform the function of hidden nodes. Moreover, this study proposes an algorithm based on the threshold accepting (TA) meta-heuristic to train the PCNN. The architecture reduces the number of weights by a great number as there are no formal connections between the input layer and the principal component layer. The efficacy of the algorithm is tested on the Spanish banks dataset and Turkish banks dataset. The results showed high generalization power of PCNN in the 10-fold cross-validation and also the feature subsets selected in each of the examples showed high discriminating power. PCNN is also compared with PCA-TANN and PCA-BPNN, which have PCA as the preprocessor and have one hidden layer each. Further comparisons are also made with TANN and BPNN. All these classifiers are compared with respect to the AUC (area under the receiver operating characteristic (ROC) curve) criterion. ROC curve is drawn for each classifier with sensitivity on the X-axis and one-specificity on the Y-axis. Based on the experiments conducted, it is inferred that the proposed PCNN hybrids outperformed other classifiers in terms of AUC. It is also observed that the proposed feature subset selection algorithm is very stable and powerful.