The Effects of Class Balance on the Training Energy Consumption of Logistic Regression Models

The presence of Artificial Intelligence and specifically Machine Learning (ML) has increased in all manner of software applications, and it already plays a major role in a variety of systems pertaining to Information Science such as public transport, disease diagnosis support and other medical problems. This increase in use has raised concerns about possible environmental impacts, since ML models require to be trained in datacentres that can impose a high ecological toll. With the aim of uncovering new ways of reducing the energy consumption of ML models, in this study we will explore the energetic impact of class balance for binary classification tasks by comparing a set of logistic regression models (LRMs) trained on a synthetic balanced dataset against another set trained on a synthetic, unbalanced dataset. We focus on the total energy and time required to complete the task, and discover that the order in energy efficiency of the models remained consistent regardless of class balance, but those trained on the unbalanced dataset required between 1.42 and 1.5 times more energy to complete the tasks, despite requiring only around 1 s more of runtime. We finish by analysing the results and proposing using synthetic datasets to estimate the energy cost of different hyperparameter options for LRMs.