Optimized fusion of spatio-temporal data for large scale construction project valuation: an attention-fused RNN-GCN model on big investment data

Construction project management requires precise, context-rich decision-making under evolving conditions, demanding models that can dynamically integrate time-dependent trends and relational structures of information. Studies have shown the potential of network-based and integrated machine learning (ML) approaches for project portfolio selection, yet advanced ML methods leveraging network-based architectures remain limited. Recently, spatio-temporal ML models have been developed that emphasize sequential forecasting and relational learning, yet they do not explicitly optimize the fusion of these perspectives, leading to suboptimal risk predictions and opaque decision processes. Motivated by this gap, this research introduces a novel SpatioTemporal Attention-based Fusion (STAF-Net), combining recurrent neural networks (RNNs) for capturing temporal dependencies and graph convolutional networks (GCNs) for modeling relational or spatial interdependencies. A newly devised attention-based fusion layer adaptively weighs these dual perspectives, allowing more explicit and accurate integration of project valuation data. Tests on a large-scale investment dataset of 11,399 construction projects demonstrate that STAF-Net surpasses simpler sequential architectures (RNN-GCN, GCN-RNN) and feature fusion RNN-GCN by consistently improving classification performance. The proposed approach attains an accuracy of 90% and an F1 score of 89%, notably outperforming benchmarks by effectively balancing the influence of temporal and spatial cues. Explicit control over the fusion of spatio-temporal data bolsters investment valuation accuracy, which is crucial for high-stakes project selection and resource allocation in construction. By incorporating a two-layer fusion mechanism, STAF-Net achieves state-of-the-art predictive accuracy, granting stakeholders enhanced representativeness and trust.