The issue of AI model bias and fairness is a long-standing issue that has recently increased interest in recent years. There are many ways to define fairness, but a common thread is that predictive systems disproportionally harm some groups. This can occur in many ways, for example, 1) when models are trained on data that is systematically biased, 2) when models make predictions that perpetuate existing societal biases, or 3) when the use of predictive systems creates or reinforces unfairness in society more broadly.
There are several ways to address AI model bias and fairness. Some methods focus on pre-processing data to remove bias, while others focus on changes to the algorithms themselves. Below we describe some specific methods for dealing with each of these issues.
1) Data pre-processing: Data pre-processing is a method of removing bias from training data. This can be done in many ways, for example, by a) sampling data points so that they are representative of the population as a whole, b) using data augmentation methods to create synthetic data points that are representative of the population or c) removing or ‘filling in’ data points that are known to be biased.
2) algorithm design: Another approach to tackling AI model bias and fairness is to design algorithms that are specifically resistant to bias. For example, decision trees can be designed to be ‘fair’ by ensuring that each decision is based on most training data points. Other algorithms have been specifically designed to be robust to different types of bias.
3) Use of domain knowledge: Another way to address AI model bias and fairness is to use domain knowledge when training predictive models. For instance, if we are building a predictive model for loan approval, we may want to use information about an applicant’s credit score, employment history, and other factors. However, we may also want to know whether the applicant belongs to a protected class (e.g., race, gender, etc.), as this information can help us avoid inadvertently biasing our model.
4) Experiments and user studies: Finally, it is essential to evaluate predictive models for bias and fairness before they are deployed in the real world. This can be done through experiments or user studies. For example, we may want to evaluate a model by looking at its predictions for different groups of people (e.g., men vs. women, different races, etc.) or by comparing its predictions to those of a ‘gold standard’ model unbiased.
The above methods are just some ways to address AI model bias and fairness. It is important to note that there is no single ‘silver bullet’ solution to this problem and that combining methods is often the most effective.
References:
https://www.kdnuggets.com/2016/09/algorithmic-bias-issues.html
https://www.oreilly.com/data/free/files/mitigating-bias-in-machine-learning.pdf
https://towardsdatascience.com/4-ways-to-prevent-algorithm-bias-in-machine-learning-2a96740fca37
https://arxiv.org/pdf/1810.03561.pdf