Filtering of inconsistent software project data for analogy-based effort estimation

Abstract

Successful software project management requires an accurate estimate for the software effort. Among the many proposed techniques, Analogy-Based Estimation (ABE) has been one of the mainstreams for software effort estimation. In general, ABE infers the effort to accomplish a new software project by reasoning on the data of the historical projects which share similar characteristics with the estimated one. As being not an exception among data mining techniques, ABE is vulnerable to the noise in the historical project data, which, consequently, needs to be filtered in advance. A number of noise filtering techniques have been proposed in both machine learning and software effort estimation domains. Basically, the effectiveness of filtering is largely dependent on the concept of noise defined by each study. Some filtering technique was claimed to have no influence on the accuracy of ABE, while some other was claimed to be vulnerable to the neighborhoods of high effort inhomogeneity. In this study, we introduce the concept of Effort-Inconsistency Degree (EID), based on which we define the inconsistent historical project data, which is the data of the historical projects that have high EID. We claim that the inconsistent historical project data were defined in a way independent of the homogeneity of effort and have high tendency to degrade the accuracy of ABE. Thus, we expect that, by filtering of these data, our approach can improve the accuracy of ABE more effectively. We have validated and compared the accuracy of ABE before and after applying our approach together with the three representative filtering techniques, namely the Edited Nearest Neighbor algorithm, the Univariate Outlier Elimination, and the Genetic Algorithm based project selection, on the four historical project data sets. The results suggest that our approach can improve the accuracy of ABE more effectively than can the other approaches.