Answers are grounded in this paper and cite their sources.
Investment decision-making in construction is influenced by a broad spectrum of uncertainties, encompassing fiscal unpredictability and geopolitical instability. Canonical valuation techniques, grounded in rational expectations, often prove insufficient in accounting for abnormal market behaviors. In response, we propose a lightweight spatio-temporal valuation architecture that incorporates graph neural network (GNN) attention modules and temporal gating operations to evaluate and classify construction projects by ex-ante investment attractiveness as defined by a decision-theoretic proxy: a structured expert-elicited multicriteria score intended to approximate early-stage portfolio screening under uncertainty (i.e., limited information at inception and time-varying macrorisk). This research utilizes a curated corpus of 12,991 global construction initiatives (downsampled from 33,684 records), supplemented with expert-formulated multicriteria rankings and Euler Hermes–derived sovereign risk ratings. Each construction project is encoded as a node within a graph, where connections are established based on regional locality and funding modality, facilitating the modeling of implicit spatial-financial dependencies. The proposed architecture family integrates a lightweight spatio-temporal attention-based graph neural network (STAGNN), which utilizes simple cosine-similarity attention mechanism for dynamic neighborhood weighting, with gated recurrent units (GRUs) to learn the temporal progression of project-level features. We formulate and evaluate multiple hybrid configurations, including different GRU variants, and nongated STAGNN variations and other standard baselines like graph convolutional network (GCN)–recurrent neural network (RNN) and graph attention network (GAT)–long short-term memory (LSTM) models in binary classification tasks (distinguishing very high from premium investment potential). Results show strong performance gains from incorporating STAGNN into spatio-temporal architectures. STAGNN reached 96% accuracy and F1 scores, recording 99% area under the receiver operating characteristic curve (AUC), underscoring the potential of a high-fidelity yet low-complexity model in minimizing false classifications and preserving high model fidelity. STAGNN’s cosine-attention eliminates dense layers and avoids multilayer perceptrons (MLPs) and multiheads, reducing parameter count and architectural complexity. It outperforms the GRU-enhanced AGNN-GAT model in prior study while keeping the design lightweight. Crucially, the attention mechanism alleviated over smoothing, preserving node-level distinctiveness in complex network topologies. The proposed framework enables rapid, data-driven investment flagging, obviating reliance on manual input, thereby advancing portfolio design strategies and risk analytics.
No takes yet. Share an insight, caveat, or question.
Mostofi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: