Identity-embedded graph neural networks for node-level cost performance index prediction in electrical construction tasks

Accurate task-level cost performance prediction in electrical construction is difficult because project data are heterogeneous, sparse, and temporally irregular. This paper proposes an Identity-Embedded Graph Neural Network (IE-GNN) model for node-level weekly cost performance index (CPI) classification. Unlike standard categorical feature encoding, the proposed method treats activity group and delivery week as identity-aware embeddings that initialize node representations before graph message passing in GNN backbones. Project progress is structured as a graph, where nodes represent tasks and edges capture their sequencing within work packages. Two high-cardinality categorical features are embedded into GNN architectures. Using 6733 electrical activities across two rolling evaluation windows, identity embeddings improved attention-based GNN performance without changing the underlying propagation mechanism. IE-GAT increased accuracy from 82%–83% to 93%–96%, while IE-TGNN increased accuracy from 82% to 92% and from 81% to 96%. The best identity-embedded models achieved 96% accuracy and 94% F1 score, supporting practical CPI risk monitoring.