Addressing Accuracy-Fairness Trade-off via Laplace Approximation Methods

The increasing deployment of machine learning models in critical domains such as finance, criminal justice, and healthcare has raised concerns about the ethical implications of these systems.

One of the primary ethical challenges is to ensure that these models are fair and unbiased. Training multiple models and choosing the best one based on fairness metrics is one of the traditional approaches for achieving fairness, which can be computationally expensive and time-consuming.

This thesis introduces a novel approach called the Laplace Approximation-Based Model Selection Framework. The framework leverages Laplace Approximation-based (LA) methods, which are employed to approximate Bayesian Neural Networks (BNNs), and its extension of Riemannian Laplace Approximation (RLA) in the context of Machine Learning model selection to address accuracy-fairness trade-off.

Leveraging LA and RLA allows us to sample multiple models from a distribution generated using a pre-trained model. Therefore, with minimal overhead, we have the option to select an optimum model from a set of models that have similar predictive accuracy scores but various fairness characteristics. We evaluated the effectiveness of the proposed framework through a series of experiments.

The results demonstrated that, to some extent, it can produce models with improved fairness scores without compromising accuracy. The experiments showed that LA and RLA approaches require 173.46 and 8.08 times less time to generate 500 models compared to the Retraining method, respectively. This thesis contributes to the development of ethical AI by exploring and offering a practical and efficient solution to the problem of model fairness with the potential of extending it to a wide range of applications.