Design Of An Iterative Deep Context Embedding And RII-Fused Evidential Framework For Predicting Construction Delay Severity Under Heterogeneous Project Environments

Abstract

Construction project schedule overruns continue to cost money, generate contractual complications, and disrupt supply chains. Decades of research have identified and ranked delay factors using indices like the Relative Importance Index (RII), but analytical and machine-learning approaches are generally limited to binary or multiclass delay classification sets. They rarely quantify prediction uncertainty, disregard contextual interactions, and treat project context as flat variables. Current models cannot generalize across regions, procurement regimes, and contractor capacities, limiting their project decision-making value. To address these limits, this work provides an end-to-end, analytically validated Deep Context Embedding and RII-Fusion pipeline for delay severity rating prediction. HPCE uses a graph-transformer architecture to learn dense contextual embeddings from heterogeneous graphs of projects, contractors, locations, and procurement trends. Format retains structural dependencies that tabular encodings lose. RII-Prior Attention Fusion (RPAF) regularizes attention weights over delay-factor embeddings using probabilistic priors to combine expert knowledge with learned context embeddings. Domain expertise is integrated into learning dynamics instead of using RII as a post hoc rating. DEOS provides a complete ordinal severity distribution, anticipated severity score, and deconstructed epistemic and aleatoric uncertainty for predictive severity modeling. This evidence-based paradigm assesses expected delay severity risk-awarely beyond point estimates. To make the model robust in various construction contexts, Counterfactual Invariant Representation Regularization (CIRR) fixes the severity mechanism across regions and procurement types and quantifies factor-level sensitivity under controlled counterfactual perturbations Finally, Conformal Prediction with Drift Guard (CPDG) ensures deployment-level reliability with calibrated prediction intervals and embedding-space drift detection for changing project conditions. The framework provides uncertainty-aware, context-sensitive severity scoring that is rigorously confirmed. The results improve forecast accuracy, adaptability across environments, and interpretability over earlier techniques. Construction delay analytics improves scheduling, contractual risk management, and policy formulations with deployable, decision-grade severity forecast.