Early Stage Fault Prediction via Inter-Project Rule Transfer

Abstract

Software fault protection can be first achieved by fault prediction. The earlier the fault prediction can be done in the software development life-cycle, the lower the damage and repair costs caused by the defects that will occur. Machine learning is one well-known method for the decision-making part of automatic software fault prediction. However, the applicability of machine learning methods is low due to the lack of data in the early stages of development processes. In this study, the data needed in the design of rule-base was obtained from counterpart projects, and the fault prediction problem was evaluated by the fuzzy rule-based systems’ point of view since these systems have portability utility which allows rule transfer between different problems with similar goals in the same domain. Briefly, this study aims to show that early-stage fault prediction is possible with the portability characteristics of fuzzy systems sourced from the inter-project rule transfer. Several experiments have been performed by using the software metrics datasets of 5 software projects to support this idea. Fuzzy systems obtained from several combinations of these datasets were evaluated by their prediction accuracy. The results show that more accurate rules can be obtained from previously completed software projects, and the use of rule bases gathered from those projects’ software metrics repositories can be transfered to predict the faulty modules of the current software project.