Title: Correcting systemic biases for fairer and more replicable data science
Presented By: Joshua Loftus, New York University

Abstract:
As use of data technologies increases it becomes more urgent for ethical data scientists to identify and correct systemic biases. In this talk I discuss two broad examples: the replication crisis in science, and the unfair and potentially illegal impacts of automated decision systems. Scientific publication suffers from selection bias, where results are more likely to be published if they appear more impressive. Algorithms in machine learning and artificial intelligence are trained on datasets that contain patterns caused by historic injustices and discrimination, and may automate the perpetuation of those injustices. Fortunately, some of these biases can be addressed with statistical methods from the emerging subfields of post-selection inference and algorithmic fairness.