Resource-Based Time and Cost Prediction in Project Networks: From Statistical Modeling to Graph Neural Networks

TL;DR

The paper tackles the problem of predicting project duration and cost under resource constraints and complex interdependencies. It introduces a Resource-Based Modeling (RBM) framework combined with Graph Neural Networks (GNNs) on a heterogeneous activity-resource graph, enriched with Bayesian online updates and active learning. Key contributions include a stochastic resource-performance model with variance propagation, a dynamic heterogeneous GNN architecture (GraphSAGE and Temporal Graph Network variants) with heteroscedastic outputs, differentiable soft critical path constraints, and online updating for adaptive forecasting; empirical results show MAE reductions of 23–31% and $R^2$ improvements from ~0.78 to ~0.91 across synthetic and benchmark datasets, along with well-calibrated uncertainty (ECE < 4% and PI coverage ~90%). The work demonstrates strong scalability, interpretability, and practical impact for risk-informed project management, enabling targeted data collection, dynamic replanning, and explainable visualizations of bottlenecks and dependencies. Future directions include incorporating causal inference, multi-fidelity modeling, and reinforcement learning to optimize decision-focused outcomes in dynamic project environments.

Abstract

Accurate prediction of project duration and cost remains one of the most challenging aspects of project management, particularly in resource-constrained and interdependent task networks. Traditional analytical techniques such as the Critical Path Method (CPM) and Program Evaluation and Review Technique (PERT) rely on simplified and often static assumptions regarding task interdependencies and resource performance. This study proposes a novel resource-based predictive framework that integrates network representations of project activities with graph neural networks (GNNs) to capture structural and contextual relationships among tasks, resources, and time-cost dynamics. The model represents the project as a heterogeneous activity-resource graph in which nodes denote activities and resources, and edges encode temporal and resource dependencies. We evaluate multiple learning paradigms, including GraphSAGE and Temporal Graph Networks, on both synthetic and benchmark project datasets. Experimental results show that the proposed GNN framework achieves an average 23 to 31 percent reduction in mean absolute error compared to traditional regression and tree-based methods, while improving the coefficient of determination R2 from approximately 0.78 to 0.91 for large and complex project networks. Furthermore, the learned embeddings provide interpretable insights into resource bottlenecks and critical dependencies, enabling more explainable and adaptive scheduling decisions.

Figures (12)

• Figure 1: Learned attention weights by neighbor type (GAT variant). Predecessors receive highest attention (0.38), followed by resource-shared activities (0.29), confirming that dependency relationships and resource interactions are most predictive.
• Figure 2: Active learning curves: RMSE on unmonitored activities vs. measurement budget. Hybrid strategy combining uncertainty and topology reduces RMSE faster than alternatives, achieving 4.61 at 60% coverage vs. 5.42 for random sampling (15% improvement).
• Figure 3: Temporal learning: prediction accuracy improves as project progresses and models incorporate execution data. TGN with Bayesian updating achieves 31% lower RMSE at 80% completion (2.73 vs. 3.97 initially) by adapting to observed resource performance.
• Figure 4: Synthetic experiments: prediction accuracy vs. project size. GNNs maintain superior performance across all scales, with larger relative gains for bigger projects (200+ activities) where neighborhood information becomes more valuable.
• Figure 5: Benchmark datasets: GNN achieves 7.8%--22.2% RMSE reduction over best non-graph baselines. Largest gains occur on PSPLIB with explicit activity networks; moderate gains on software datasets with coarser structure.