Toward Accurate Software Effort Prediction Using Multiple Classifier Systems

Abstract

Averaging is a standard technique in applied machine learning for combining multiple classifiers to achieve greater accuracy. Such accuracy could be useful in software effort estimation which is an important part of software process management. To investigate the use of ensemble multiple classifiers learning in terms of predicting software effort. The use of ensemble multiple classier combination is demonstrated and evaluated against individual classifiers using 10 industrial datasets in terms of the smoothed error rate. Experimental results show that multiple classifier combination can improve software effort prediction with boosting, bagging and feature selection achieving higher accuracy rates. Accordingly, good performance is consistently derived from static parallel systems while dynamic classifier selection systems exhibit poor accuracy rates. Most of the base classifiers are highly competitive with each other. The success of each method appears to depend on the underlying characteristics of each of the ten industrial datasets.