What question did this study set out to answer?

The aim is to improve the standardization of Bill of Quantities through automated classification of Work Classification Codes.

April 21, 2026Open Access

Hierarchical Ensemble Learning for Automatic Classification of Construction Work Codes

Puntos clave

The aim is to improve the standardization of Bill of Quantities through automated classification of Work Classification Codes.
Constructed a dataset of 8,021 items from school construction BOQs.
Developed an automated matching model using hierarchical ensemble learning with distinct methods for varying complexity levels.
Conducted baseline comparisons with a single LSTM model to evaluate performance differences.
Ensemble methods outperformed the baseline LSTM model, especially at intermediate classification levels.
Bagging improved stability at upper levels, while Boosting showed significant gains at intermediate levels.
Level 5 classification experienced performance decline due to data sparsity issues.

Resumen

The Work Classification Code (WCC) serves as a critical tool for standardizing Bill of Quantities (BOQ) in construction projects, ensuring reliable cost estimation and effective expense management. However, current BOQ–WCC matching processes rely heavily on subjective judgment, resulting in inconsistencies and limited standardization. To address this issue, this study constructed a dataset of 8,021 valid items collected from six school construction BOQs commissioned by the Korea Public Procurement Service (PPS), and developed an automated matching model using hierarchical ensemble learning. The proposed approach reflects the five-level hierarchical structure of the WCC (Levels 1–5) and applies different ensemble methods according to classification complexity: Bagging at Levels 1–2, Boosting at Levels 3–4, and Weighted Voting at Level 5. A baseline comparison with a single LSTM model was also conducted, confirming that ensemble methods consistently outperformed the base model, particularly at intermediate levels where classification complexity is higher. Experimental results demonstrated that Bagging provided stable improvements at upper levels, Boosting achieved substantial gains at intermediate levels, while Weighted Voting offered limited benefits at the lowest level. Despite these improvements, overall performance declined as class granularity increased, and Level 5 classification showed severe degradation due to extreme data sparsity. This study confirms the feasibility of applying hierarchical ensemble learning for automated WCC matching and highlights its potential for improving BOQ standardization in real-world projects. The findings suggest that automatic WCC assignment can support reliable cost reviews and enable retroactive coding of BOQs where codes were previously absent, thereby enhancing efficiency in design-phase verification systems. Future work will focus on addressing data sparsity through large-scale BOQ collection, incorporating advanced pre-trained language models such as BERT, refining the hierarchical code structure through data integration, and validating model applicability in public procurement and contractor systems.

Me gusta

Guardar

Ver artículo completo