(An) empirical analysis of software effort estimation with outlier elimination

Abstract

Accurate estimation has always been challenge for software engineering communities. Many researches have done studies where estimation models were compared to choose the best accurate model or new estimation models were proposed to improve the prediction accuracy. However, many works did not consider the data set which we believe is a basis to build accurate estimation model. The data set often has a faulty, incomplete data and extreme value data. Such data is called an outlier. Therefore, the outlier need to be handled to build a better model. In this thesis, we investigate the prediction accuracy of effort estimation models when applying outlier elimination techniques. Three commonly used effort estimation models, and two outlier elimination techniques are selected for our empirical study. The empirical results show that the prediction accuracy of effort estimation models with outlier elimination techniques are more accurate than that of effort estimation models which is not applied the outlier elimination techniques. In addition, our study shows different result depend on the models based on two different data samples. Our study can be used in organizations to build effort estimation model for current or future projects.