Machine learning and data analysis have enjoyed tremendous success in a broad range of domains. These advances hold the promise of great benefits to individuals, organizations, and society as a whole. Undeniably, algorithms are informing decisions that affect all aspects of life, from news article recommendations to criminal sentencing decisions to healthcare diagnostics. This progress, however, raises (and is impeded by) a host of concerns regarding the societal impact of computation. In a short but intensive cluster at the Simons Institute, we discussed the concern of algorithmic fairness: will algorithms discriminate or make more equitable decisions? Unfortunately, we can consistently observe that, left to their own devices, algorithms may propagate and possibly even amplify existing biases. Addressing discrimination by algorithms (as well as other societal concerns) is not only mandated by law and ethics, but is essential to maintaining the public trust in the computation-driven revolution we are experiencing.
The study of fairness is not new, and is truly multidisciplinary: philosophers, legal experts, economists, statisticians, social scientists and others have been concerned with fairness for generations. Nevertheless, the scale of decision-making in the age of Big Data, as well as the computational complexities of algorithmic decision-making, implies that computer scientists must take an active part in this research endeavor. Indeed, computer scientists are rising to the challenge, as manifested by an explosion of research in recent years. Many of the approaches discussed in our cluster, while influenced by other disciplines, also inherently rely on computer science insights. The theory of computing (TOC) community was early within computer science to identify the importance as well as the intellectual appeal of algorithmic fairness, and early theory works have proven to be quite influential. In this cluster, we discussed recent works as well as many exciting avenues for future research. The picture that unfolded is of a rich and challenging area of research that is ripe for contributions from a wide variety of theory subareas. In the following, we will mention a small collection of the discussed topics at a very high level.
Past successes of TOC, such as cryptography, game theory and differential privacy, have demonstrated the importance of introducing the right models and the right definitions. Models and definition for algorithmic fairness have been the topic of much research in recent years, and occupied a considerable part of our discussion in the cluster. For example, we discussed statistical notions of fairness pertaining to large populations in comparison with individual notions, as well as promising definitions that lie in between. We noted that algorithmic fairness is subtle: natural definitions can be abused and some stand in contradiction to each other. In addition, one shouldn't expect a single definition of algorithmic fairness to apply to all contexts and to all societies. Having said that, versatile models and powerful definitions have been introduced and studied, leading both to novel algorithms as well as instructive lower bounds. Some of the questions we studied are:
- How can unfairness be identified and mitigated in each step of the algorithmic pipeline (in data collection and processing as well as data analysis)?
- To what extent is fairness aligned with accuracy and utility, and to what extent are the two in conflict?
- What are potential (undesirable) interactions between computational components with respect to the fairness of the larger systems they compose?
- Can effective notions of fairness be based on the rich literature of causality? Can there be meaningful fairness guarantees without complete understanding of the relevant causal structure?
- What are the impacts of algorithmic choices in terms of long-term fairness?
Each of these questions relates to promising recent works, and suggests many exciting directions for future research. If nothing else, one should read this short summary as a passionate call for contributions to the understanding of algorithmic fairness and the societal impact of computation more generally. This field needs the techniques of TOC as well as its famous optimism in the face of complexities, as part of a multidisciplinary and holistic effort.
A larger cluster on this topic will be held at the Simons Institute in Summer 2019. More information is forthcoming shortly.