MACHINE LEARNING FOR PREDICTING SOFTWARE PROJECT FAILURE RISKS

Abstract

Software engineering industry still face failure of software projects which result in large losses, waste of resources and poor software quality. Existing risk assessment techniques use static, expert based heuristics, which do not adapt to dynamic development cycles of modern software development. This research presents a machine learning driven predictive framework that uses heavy header engineering in multiple dimensions, including software complexity metrics, project management indicators, human factors, defect trend and financial constrains to assess and mitigate software project failure risks. Using the above algorithm, the proposed approach takes advantages of various supervised learning models, like Random Forest, Gradient Boosting Machines (GBM), etc. and unsupervised anomaly detection models to discover hidden risk patterns. Then, we evaluate model effectiveness with precision recall analysis, F1 score and Matthews Correlation Coefficient (MCC) on real world software project dataset. Results indicate that our model characterizes leak early failure prediction with high accuracy, better than others risk assessment methods. Moreover, we use SHAP and LIME to answer representation and interpretability questions in software managers. This work provides a backbone to implement AI driven risk prediction in an Agile software development to make proactive decisions and avoid project failure before they grow.