Cost-sensitive And Ensemble-based Intelligent Prediction Model For Outsourced Software Project Risk

Outsourcing is one of the major ways to develop software today and is associatedwith a high failure rate. In order to improve the success ratio in software development,much effort has been invested in creating intelligent prediction models for riskmanagement. The objective of such models is to assist software managers inevaluating success ratio and cost-effectiveness.This study has summarized researches of intelligent prediction models forsoftware project risk over the past20years and found that there are few researchesintroduce cost-sensitive learning and ensemble method into this field to the best of ourknowledge. All of the existing prediction models, however, are based on thehypothesis that the misclassification costs are equal. It does not reflect the realsituation in software risk management where the inability to recognize project failureis far more serious than misclassifying a successful project as a failure. To considerthis fact, this study aims to explore the cost-sensitive prediction model for softwareproject risk prediction. Moreover, to enhance prediction performance and lower itscost, the research explores homogenous and heterogeneous ensemble machinelearning methods and compares them to common algorithms. We also applied tests ofstatistical significance (T-test) to evaluate whether the superior performance of onemodel over another is statistically significant.Based on a collected sample of327outsourced software projects, weinvestigated the performance of60different models in combinations or as singlemodels. The experimental results suggest that the best choice is to build cost-sensitivesoftware risk prediction models with bagging-based homogenous decision treesmodels. The decision tree is appropriate to software project practice due to its goodprediction accuracy, and it is simple and easy to explain.This research is one of the first to use cost-sensitive misclassification parameterand ensemble method to construct the software risk management prediction modeland develops a more comprehensive method for prediction model evaluation. Inaddition, this research provides practitioners with a clear framework on cost-sensitiveand ensemble based prediction modeling of outsourced software project risk.