Carolyn Mair and Martin Shepperd
The rationale for this EPSRC funded project, Meta-level Learning for Software Project Prediction (MeLLow), was the existence and generation of empirical studies to evaluate the accuracy of project effort prediction systems, but with inconsistent results. Thus the aim of the project was to combine results from these primary empirical studies and use them as training data for meta-level learners to induce decision trees to advise practitioners which prediction system should be used in what context. The main contributions of this research were to have identified problems (methodological and reporting) that inhibit meaningful meta-analysis, conducted systematic reviews, and proposed a reporting protocol for future primary studies. In addition we have carried out further data analysis for our industrial partners, STFF, Finland and BT, UK, and explored the use of sophisticated modelling techniques for software productivity, which is a key component of effort prediction.
Key words: software engineering, software project estimation, data analysis, meta-analysis, systematic review.