Abstract/Summary

ASHRAE Global Thermal Comfort Database has been extensively used for analyzing specific thermal comfort parameters or models, evaluating subjective metrics, and integrating with machine learning algorithms. Outlier detection is regarded as an essential step in data preprocessing, but current publications related to this database paid less attention to the influence of outliers in raw datasets. This study aims to investigate the filter performance of different outlier detection methods. Three stochastic-based approaches have been performed and analyzed based on the example of predicting thermal preference using the Support Vector Machine (SVM) algorithm as a case study to compare the predictions before and after outlier removal. Results show that all three rules can filter some obvious outliers, and the Boxplot rule produces the most moderate filer results, whereas the 3-Sigma rule sometimes fails to detect outliers and the Hampel rule may provide an aggressive solution that causes a false alarm. It has also been discovered that a small reduction in establishing machine learning models can result in less complicated and smoother decision boundaries, which has the potential to provide more energy-efficient and conflict-free solutions.