Implications Tutorial - Reasoning About Subtle Biases in Data

Explore the implications of subtle biases in data and their impact on the reliability of decision support tools in this comprehensive tutorial from the FAT* 2019 conference. Delve into the challenges of deploying machine learning-driven decision aids, focusing on how common assumptions in model training often fail to hold in real-world applications. Examine case studies demonstrating dangerous predictions resulting from biased training data, and learn to identify selection bias and other factors that can compromise model generalization. Gain insights into detecting less obvious biases beyond protected attributes like race or gender, and understand how the deployment of decision support tools can itself lead to shifts in practice and policy. Acquire concepts and terminology to frame issues related to dataset shift and apply them to your own applications. Conclude with an overview of available solutions, their applicability, and their respective advantages and disadvantages in addressing data bias challenges.