Summary contributed by our Artist-in-Residence Falaah Arif Khan. She’s also a Research Fellow in the CVIT Lab at the International Institute of Information Technology.
Link to original paper + authors at the bottom.
Mini-summary: In this succinct review of the scholarship on Fair Machine Learning(ML), Chouldechova and Roth outline the major strides taken towards understanding algorithmic bias, discuss the merits and shortcomings of proposed approaches, and present salient open questions on the frontiers of Fair ML. These include- statistical vs individual notions of Fairness, the dynamics of fairness in socio-technical systems, and the detection and correction of algorithmic bias.
The motivation behind the paper is to highlight the key research directions in Fair ML that provide a scientific foundation for understanding algorithmic bias. These broadly include- identifying bias encoded in data without access to outcomes (for example we have access to data about who was arrested and not who committed the crime), the utilitarian approach to optimization and how it caters purely to the majority without taking into account minority groups and the ethics of exploration. The role of exploration is a key one since in order to validate our predictions we must have data that enumerates how the outcome in fact played out. This brings up several important questions such as: Is the impact of exploration overwhelmingly felt by one subgroup? If we deem the risks of exploration too high, by how much does a lack of exploration slow learning? Is it ethical to sacrifice the well-being of current populations for the perceived well-being of future populations?
The next important research direction is one that seeks to formalize the definition of Fairness. There are several proposed definitions, the most popular one being the statistical definition of Fairness. Such a formulation enforces parity in some chosen statistical measure across all groups in the data. The simplicity, assumption-free nature, and the ease with which a statically fair allocation can be verified makes this definition popular. However, a major shortcoming is the proven impossibility of simultaneously equalizing multiple desirable statistical measures. A statistical definition of fairness can also be computationally expensive to model.
The second popular notion is that of Individual Fairness, which enforces that, for a given task, the algorithm treats individuals who are similar, similarly. While this is richer, semantically, it makes strong assumptions that are difficult to realize practically.
Chouldechova and Roth then go on to present questions around Intersectional Fairness, namely: how different algorithmic biases compound for individuals who fall at the intersection of multiple protected groups. They also question the feasibility of a ‘good’ metric of fairness and whether such a notion will be accessible while making predictions, and the existence of an ‘agnostic’ notion of Fairness that does not rely on any one measure, but instead takes human feedback to correct for bias.
Another important consideration is the dynamics of Fairness. Models are seldom deployed in one-shot settings and are usually used in conjunction with several other predictors. In such a setting, how does compositionality affect algorithmic fairness? ie. do individual components that satisfy conditions of ‘fairness’, continue to adhere to the same degree of fairness when composed together to decide a single outcome?
Another source of dynamism is the impact that algorithmic decision-making systems have on the environment. Models that determine outcomes also influence the incentives of those who interact with them and hence it becomes imperative to consider long-term dynamics when designing ‘fair’ algorithms. We also need to reconcile the individual motives of the different actors in the system and incentivize them to behave ethically.
Lastly, Chouldechova and Roth enumerate open questions in modeling and correcting for bias in data, namely: How does bias arise in data? How do we correct for it? How do we take into account feedback loops, where biased predictions further lead to biased training data in future epochs? Enforcing any notion of fairness on biased data would see a drop in model accuracy and this begs the question of how we go about validating our ‘fair’ predictions.
Original paper by Alexandra Chouldechova and Aaron Roth: https://dl.acm.org/doi/pdf/10.1145/3376898