“Fairness” and “Machine Learning” both mean many things, which naturally complicates the study of fairness in machine learning. Hence, it is important to specialize. In this workshop, we will focus on two fair learning paradigms: “interactive” methods, and representation learning. This workshop aims to identify key challenges and open questions that currently limit both our theoretical understanding of these methods, and their applicability in practice.

Interactive machine learning methods operate in settings in which the actions of the learning algorithm feedback into its data collection, or change the environment in which it is operating, or both. This class of settings includes both bandit optimization and reinforcement learning (among others), and is distinct from batch classification.

Interactive learning settings naturally arise in many fairness-relevant learning tasks: Predictive policing algorithms are able to observe crime to higher fidelity in areas to which they have deployed officers (and may also reduce crime in areas to which more police officers have been deployed). Lending and criminal recidivism prediction algorithms, when actually used to make decisions, observe outcomes corresponding to only one of their decisions (i.e. loan payback only if the loan was granted, recidivism only if the individual was released). The feedback loops that can result when this structure has not been properly taken into account has been blamed as one source of unfairness. However, “exploration”, which is necessary to avoid these feedback loops, has also been identified as a potentially unfair operation, when the decisions of the algorithm correspond to important actions affecting the lives of individuals. How should we think about managing these tradeoffs when fairness concerns predominate?

Fair representation learning is the process of constructing transformations of the original data that retain as much of the task-relevant information as possible while removing information about sensitive or protected attributes. Informally, these may be thought of as a form of data “de-biasing”. This shifts the fairness burden away from the particular choice of prediction method towards identifying data transformations that can be used in combination with any desired modeling approach. For instance, by removing all information that can be used to infer a defendant’s race from a set of criminal data, one can construct recidivism prediction models that are in a sense “race-neutral”. Yet questions remain about what kinds of fairness such representations can effectively promote, and whether we risk introducing any other types of biases along the way. How fair are fair representations, really?