Cost-Duration Variance Prediction (CDVP) System Using Explainable Bayesian Optimized Ensemble Machine Learning

Abstract

Abstract This research introduces an integrated system for a novel management concept called automated cost-duration variance prediction (CDVP) system which allows to the project managers to look ahead at project budget, duration, and its variances at the project completion. To obtain the optimal results of each algorithm, the proposed system conducted Bayesian optimization for each developed algorithm. As a result, Extremely Randomized Trees (ET) was the optimal model for duration prediction at 2.53% and 0.961 for Mean Absolute Percentage Error (MAPE) and adjusted determination coefficient (R*2), respectively. AdaBoost Regressor (ABR) was the best model for duration variance prediction with 8.970 MAPE, and 0.949 R*2. On the other hand, the optimal model for project cost was a Light Gradient Boosting Machine (LGBM) with 2.98% and 0.931 for MAPE and R*2, respectively. Extremely Randomized Tree (ET) was the best model for cost variance prediction with 4.151 MAPE, and 0.962 R*2. SHapley Additive exPlanations (SHAP) technique was employed to provide explanations for the key drivers that used as inputs for both cost and duration predictions.