In-Processing fairness improvement methods for regression Data-Driven building Models: Achieving uniform energy prediction

Abstract

In recent years, the massive data collection in buildings has paved the way for the development of accurate data-driven building models (DDBMs) for various applications. However, a model with a high overall accuracy would not ensure a good predictive performance on all conditions. The biased predictive performance for some conditions may cause fairness problems. Although pre-processing methods were proposed to improve predictive fairness by removing discrimination from training datasets for classification problems in building engineering domain, they lack the ability of achieving user-defined trade-off between fairness and accuracy for regression problems, such as energy prediction. To improve the predictive fairness of regression models in terms of having similar predictive performance between different conditions, this study proposes four in-processing methods, namely mean residual difference penalized (MRDP) regression, mean square error penalized (MSEP) regression, mean residual difference constrained (MRDC) regression, and mean square error constrained (MSEC) regression, to add fairness-related penalties or constraints to the loss function of regression models. Then, these proposed methods are applied to develop linear regression models for energy prediction of an apartment. In this case study, improving predictive fairness means to let the energy predictive accuracy be uniform no matter if there is occupancy movement. The result shows that MSEC is the most powerful method to improve fairness in terms of mean square error (MSE) rate and mean absolute error (MAE) rate, while MSEP is another good option to improve fairness without a significant decrease on the overall accuracy. MRDC is effective on improving the similarity of absolute mean residual difference (abs(MRD)) between different conditions, however, MRDP would not affect the predictive result.