Explainable machine learning for project management control

Abstract

Project control is a crucial phase within project management aimed at ensuring —in an integrated manner— that the project objectives are met according to plan. Earned Value Management —along with its various refinements— is the most popular and widespread method for top-down project control. For project control under uncertainty, Monte Carlo simulation and statistical/machine learning models extend the earned value framework by allowing the analysis of deviations, expected times and costs during project progress. Recent advances in explainable machine learning, in particular attribution methods based on Shapley values, can be used to link project control to activity properties, facilitating the interpretation of interrelations between activity characteristics and control objectives. This work proposes a new methodology that adds an explainability layer based on SHAP —Shapley Additive exPlanations— to different machine learning models fitted to Monte Carlo simulations of the project network during tracking control points. Specifically, our method allows for both prospective and retrospective analyses, which have different utilities: forward analysis helps to identify key relationships between the different tasks and the desired outcomes, thus being useful to make execution/replanning decisions; and backward analysis serves to identify the causes of project status during project progress. Furthermore, this method is general, model-agnostic and provides quantifiable and easily interpretable information, hence constituting a valuable tool for project control in uncertain environments.