Contextualizing the Accuracy-Fairness Tradeoff in Algorithmic Prediction Outcomes

Abstract

In this paper, a probabilistic machine learning method is proposed to predict the indoor temperature of an office environment. An IOHMM-based model is developed to represent the office environment under different circumstances of heating sources. One year of time series data is observed and studied to learn the dynamics of the indoor thermal states. The uncertainty associated with the changing aspects of the indoor temperature and its dependence on the outdoor temperature is considered in the model development. The well-known Baum Welch and forward-backward algorithms are adapted to learn the model parameters. Then, the Viterbi algorithm is used to predict the maximum path of hidden states, leading to predicting the most likely future temperatures. A numerical application is presented to demonstrate the model development steps and the training and testing results. Finally, the model's performance is validated using leave-one-out cross-validation, which shows that the model has a prediction accuracy of about 78%.