Software Effort Classification with Multilayer Perceptron Neural Networks

Abstract

We target the problem of software effort estimation from a classification perspective.Our main goal is to build a classifier that can predict the required effort for a new project and assign it into one of the four classes: small, small-medium, medium-large, and large.A criterion is proposed for the aforementioned categorization and based on the amount of required effort.We study data sets that are based on three different categories: function points-based data sets, COCOMO-based data sets, and project characteristics-based data sets.Feature sets are prepared and fed to the multilayer perceptron (MLP) neural network algorithm.A hold-out test is implemented and ROC curve is used as a measure of the performance of the algorithm.In addition, we identify the important features for building the classification models across various data sets.Generally, MLP shows good performance across the six data sets.Moreover, there is a stability of performance within each category of data sets or across different categories with no dramatic differences in results.However, the performance of MLP seems to decrease with data sets that are based on project characteristics.