DATA ENRICHMENT IN CASE-BASED REASONING FOR SOFTWARE COST PREDICTION

Key Results: Cases are often indiscriminatingly added to case bases. This potentially results in poor solutions proposed by ‘misleading ’ cases. Research to-date has resulted in development of a prototype that can potentially filter ‘misleading ’ cases in case bases to avoid delivery of poor solutions. The prototype was tested in a pilot study, in which absolute residuals (error in prediction) were reduced by 45.2 % from 68.17 hours to 37.33 hours. How does the work advance the state-of-the-art? Our research explores methods to enrich data sets that use feature vectors for data representation. We identify the limitations of such representation and propose two methods to address deficiencies. Though the research is specific to software engineering data sets, the proposed techniques can be easily adapted to suit other problem domains that use similar forms of data representation. Motivation (Problems addressed): Feature vectors pose constraints on the depth and richness of information about the cases in case bases. The lack of comprehensive knowledge leads to application of simplistic adaptation routines for case modification in case-based reasoning systems. Hence, proposed solutions are sub-optimal. This is a generic problem that can be observed in data sets across a variety of domains