Feature Selection Technique for Effective Software Effort Estimation Using Multi-Layer Perceptrons

The software effort estimation is essentially required to be done effectively and accurately for delivering quality product within the budget limits on time. A feature selection technique is proposed to effectively estimate effort using non-linear model based on the multilayer perceptron architecture. The objective is to find out whether the feature selection technique improves the accuracy of the prediction model developed. A prediction model (PRED_MLP) is built using Multilayer perceptron architecture with back propagation algorithm. Accuracy of the proposed model is compared with the accuracy of another proposed model (PRED_MLP_FS) which is backed with feature selection technique based on neighborhood component analysis. The dataset from Desharnais Project are used. The accuracy of proposed models is assessed and empirical comparison is also made between the prediction powers of these two predictors using standard metrics. Both the proposed models namely PRED_MLP and PRED_MLP_FS are validated. The experimental work shows that the model PRED_MLP_FS outperforms the model PRED_MLP. The results are statistically significant and suggest that the feature selection techniques can improve the accuracy of the prediction model upto 40%. Therefore, some input parameters can be dropped without loss in estimation accuracy.