A recommendation-based directed-graph framework for construction activity sequencing
APPLIED SOFT COMPUTING JOURNAL, cilt.202, ss.1-21, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 202
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.asoc.2026.115937
- Dergi Adı: APPLIED SOFT COMPUTING JOURNAL
- Derginin Tarandığı İndeksler: Applied Science & Technology Source, Scopus, Science Citation Index Expanded (SCI-EXPANDED), Compendex, INSPEC
- Sayfa Sayıları: ss.1-21
- Karadeniz Teknik Üniversitesi Adresli: Evet
Özet
Despite increased digitization, the underlying logic of construction sequencing often lacks an explicit structure for data-driven successor-activity recommendation. Existing models, particularly classification-based formulations, may suffer from performance degradation under data sparsity and class imbalance, limiting their operational utility. Furthermore, deep sequence models may be difficult to apply effectively in construction contexts because they typically benefit from dense, repetitive sequences and feature-rich activity descriptions. To address this gap, this study develops a recommendation-based directed-graph framework that formulates construction activity sequencing as a link-prediction problem. Standardized construction activities are represented as nodes, while observed predecessor–successor relationships are represented as directed edges, enabling plausible successor activities to be ranked from historical schedule structures. The framework evaluates Association Rule Mining (ARM) as a rule-based recommendation method, DeepWalk and node2vec as random-walk embedding methods, and GCN, GAT and GraphSAGE as graph neural network (GNN) benchmarks. Experiments were conducted on directed activity graphs derived from 24 completed construction-project schedules across six test groups with different network scales. ARM achieved approximately 78% accuracy in the smaller test group, where recurrent sequencing relations were more evident. In the larger test groups, DeepWalk achieved approximately 75% accuracy, indicating stronger generalization as graph size and structural diversity increased, whereas the GNN baselines showed lower and less stable results. In the practical evaluation of Project P9, ARM achieved 77% accuracy, generated a sequence containing 46 activities and flagged 13 activities for manual review. The findings demonstrate that framing construction activity sequencing as a recommendation-oriented link-prediction problem enables scalable and data-efficient generation of editable, project-specific draft schedules while preserving planner oversight under sparse construction graph conditions.