What needs to be done to make a fair algorithm?

Question

What needs to be done to make a fair algorithm (supervised and unsupervised)?

In this context, there is no consensus on the definition of fairness, so you can use the definition you find most appropriate.

Answer 1

The paper Fair and Unbiased Algorithmic Decision Making: Current State and Future Challenges argues that ensuring fairness is not a trivial task and that the current statistical formalizations of fairness lead to a long list of criteria that are each flawed (or even harmful) in different contexts, that is, there are trade-offs between the proposed formalizations. Therefore, fairness constraints in algorithms have to be specific to the domains to which the algorithms are applied. To achieve that, there is a need for collaboration with domain experts and explainable artificial intelligence.

The major obstacle towards fair machine learning algorithms is the presence of algorithmic bias, which can be subdivided into the following main categories:

• the bias in the data, and
• the inductive bias (the implicit or explicit assumptions behind the algorithm or model) or, in general, any bias introduced during the development of the algorithm or model (for example, a certain choice of a subset of features can change the outcome of the model).

The bias in the data can be due to different factors, such as a biased choice of the collected (or labelled) data or measurement errors (which can make the data not representative of the population). Causal inference can be used to understand the causal relationships in the data, thus it can be used to find the source of bias in the data. To avoid unfairness due to bias in the data, there is a need to analyze and understand the data, so that to improve its quality (for example, by increasing the diversity of the data). However, the bias in the data is not always easily reducible, given that certain outcomes of an experiment may rarely occur or are hard to produce in practice, so unbiased data may not always be easily collectible.

In the paper The selective labels problem: Evaluating algorithmic predictions in the presence of unobservables (2017), the authors address the selective labels problem (that is, the judgments of decision-makers determine which instances are labeled in the data, which can thus introduce bias in the data) and develop an approach called contraction, which can be used to compare the performance of predictive models and human decision-makers (even in the presence of unobservables). There are also works that are based on Bayesian or causal inference (for example, risk-adjusted regression)

The sample bias (the data is not representative of the overall population due to a systematic intentional or unintentional error in data collection, a measurement error, which can also be due to social prejudices) is also a form of bias in the data. In the paper Residual unfairness in fair machine learning from prejudiced data (2018), the authors address this problem in the context of and of police stop-and-frisk (where a biased police behavior leads to over-proportional stopping of a racial minority group). Nathan Kallus and Angela Zhou show that adjusting the classifier for fairness does not solve the sample bias problem.

In the paper Fair and Unbiased Algorithmic Decision Making: Current State and Future Challenges, the author argues that any attempt to reduce the bias introduced during the development of the algorithms or models, if it does not take into account the specific social and moral context where they are supposed to be applied, can still lead to algorithmic bias. Furthermore, algorithms should undergo frequent reevaluations, given that, for example, the underlying population or the application context may change.