Comparative evaluation of data mining methods in predicting the water vapor permeability of cement-based materials

Water vapor permeability of building materials is a crucial parameter for analysing and optimizing the hygrothermal performance of building envelopes and built environments. Its measurement is accurate but time-consuming, while data mining methods have the potential to predict water vapor permeability efficiently. In this study, six data mining methods—support vector regression (SVR), decision tree regression (DT), random forest regression (RF), K-nearest neighbor (KNN), multi-layer perceptron (MLP), and adaptive boosting regression (AdaBoost)—were compared to predict the water vapor permeability of cement-based materials. A total of 143 datasets of material properties were collected to build prediction models, and five materials were experimentally determined for model validation. The results show that RF has excellent generalization, stability, and precision. AdaBoost has great generalization and precision, only slightly inferior to the former, and its stability is excellent. DT has good precision and acceptable generalization, but its stability is poor. SVR and KNN have superior stability, but their generalization and precision are inadequate. MLP lacks generalization, and its stability and precision are unacceptable. In short, RF has the best comprehensive performance, demonstrated by a limited prediction deviation of 26.3% from the experimental results, better than AdaBoost (38.0%) and DT (38.3%) and far better than other remaining methods. It is also found that data mining methods provide better predictions when cement-based materials’ water vapor permeability is high.