Applying Cross Project Defect Prediction Approaches to Cross-Company Effort Estimation

Abstract

BACKGROUND: Prediction systems in software engineering often suffer from the shortage of suitable data within a project. A promising solution is transfer learning that utilizes data from outside the project. Many transfer learning approaches have been proposed for defect prediction known as cross-project defect prediction (CPDP). In contrast, a few approaches have been proposed for software effort estimation known as cross-company software effort estimation (CCSEE). Both CCSEE and CPDP are engaged in a similar problem, and a few CPDP approaches are applicable as CCSEE in actual. It is thus beneficial for improving CCSEE performance to examine how well CPDP approaches can perform as CCSEE approaches. AIMS: To explore how well CPDP approaches work as CCSEE approaches. METHOD: An empirical experiment was conducted for evaluating the performance of CPDP approaches in CCSEE. We examined 7 CPDP approaches which were selected due to the easiness of application. Those approaches were applied to 8 data sets, each of which consists of a few subsets from different domains. The estimation results were evaluated with a common performance measure called SA. RESULTS: there were several CPDP approaches which could improve the estimation accuracy though the degree of improvement was not large. CONCLUSIONS: A straight forward application of selected CPDP approaches did not bring a clear effect. CCSEE may need specific transfer learning approaches for more improvement.