DSCostPred: a double-stacking model for construction cost prediction

The prediction of construction project cost plays a core role in engineering construction projects. However, the current prediction involves a multi-dimensional and dynamically variable system, and each major category can be further subdivided into many specific factors. Meanwhile, variables' relationships present a complex network of nonlinearity and interaction, which seriously affected the prediction accuracy. To solve this problem, we proposed a dual-stacking construction cost prediction method based on variable stacking and model stacking (DSCostPred). This method emphasizes that classifying variables and applying different algorithms respectively can avoid the impact of variables' functional differences. First, the variables are pre-classified to avoid mutual interference among them. Then, to learn the attribute and function positioning, as well as the complex interaction among them, different types of models are utilized to learn the variables. In algorithm design, to achieve the organic combination of multiple attributes and multiple models, a variable stacking is introduced into stacking ensemble learning to form collaborative predictions with model stacking. This method was compared with the classical method on real data, and the results show the superior performance. In addition, the ablation experiments and SHAP analysis also demonstrated the feasibility of the double-stacking idea we proposed.