Multi-Objective and Multiple Information Source Optimization for Fair & Green Machine Learning

Democratizing Machine Learning (ML) requires endowing ML algorithms with fairness and low environmental impact. Training ML models on real-life data might lead to inherently biased predictions, a critical issue when the bias translates into discrimination of certain social groups. This unfairness is exacerbated by searching for the best predictive model only depending on accuracy. FairML was initially addressed as a constrained optimization problem, but recently multi-objective methods proved to be more effective. The second issue, considered here, is the carbon footprint of ML: the massive usage of computational resources for training ML models, and searching for the best one, implies a significant environment impact, leading to Green AutoML methods. Recent approaches estimate the carbon footprint as a proxy of the energy consumption and reduce it by using multiple information sources (i.e., small portions of the data), each with a different fidelity and cost. We propose a method combining multi-objective and multiple information source into a single Bayesian optimization framework. It was evaluated on four fairness datasets and three ML algorithms, and compared those from state-of-the-art methods. Results empirically prove that our method outperforms the others in terms of accuracy, fairness, and “greenness”.