Use of the circle segments visualization technique for neural network feature selection and analysis

In this paper, the circle segments (CS) technique is proposed as a data visualization tool for selecting and analysing the effects of the input features towards the target outputs in constructing neural network models. Specifically, the multi-layer perceptron (MLP) network is employed to tackle function approximation and pattern classification tasks, and CS is used to provide visualization of the correlations between the input features and the target outputs in these tasks. The effectiveness of the proposed approach is evaluated using two benchmark data sets, one for function approximation and another for pattern classification. Performance comparison with the response surface methodology (in function approximation) and with principal component analysis (in pattern classification) is conducted. The results indicate the usefulness of CS in examining the correlations between the input–output data samples, with improved performances. In addition, a real medical diagnosis task is used to evaluate the applicability of the approach. The outcomes, again, demonstrate improvements in accuracy, sensitivity, and specificity with the use of CS for feature selection, even with more than 50% of the input features eliminated.