Authors: S. Balaji, N. Poyyamozhi
Abstract: Currently, the field of project management faces increasing uncertainty as projects must deal with changing requirements, resource shortages, and the unpredictable effects of human actions, technical systems, and external events. However, existing data-driven models have failed to provide interpretable results, preventing project managers from identifying the factors that create risks. Thus, this research presents a lightweight and explainable data-driven decision support system that enables project risk prediction and risk management in complex project management environments. The devised methodology employs a Project Management Risk Dataset, which includes project demographics and operational metrics, human factors, organizational context, technical aspects, and external influences. Moreover, a comprehensive data reliability testing is conducted through pre-processing methods for categorical attributes, one-hot encoding, and Min-Max normalization of budget and timeline, and risk metrics. Advanced feature engineering uses graph-based feature relationships to identify hidden project attribute dependencies, Graph Signal Processing to create project attribute dependencies, and LASSO with polynomial feature expansion to achieve optimal results. The proposed TAM-Lite architecture integrates TabNet, a mini autoencoder, and a shallow multilayer for project risk prediction. Moreover, stage-wise training is conducted based on Gradient Boosted Rule Sets with Extreme Learning Machines and fuzzy logic classification. The model generates risk level probabilities, which are evaluated through Bayesian Networks and counterfactual explanations to deliver clear and actionable risk reduction recommendations.
DOI: https://doi.org/10.5281/zenodo.19483001
