Author(s), Ed Shee Machine Learning. Photo by Lukas at Unsplash. Over the past few years there has been an emphasis on creating machine learning systems that are responsible and ethical. Although the terms “Bias”, Fairness, and Explainability are frequently used in the machine learning world, their meanings can be confusing and often misunderstood as the same thing. We’re going to try to clarify that. Bias. Before we get into machine learning and how it affects our lives, let us first look at the definition of bias in the dictionary. “Inclination or prejudice against or for one individual or group, particularly in a manner considered unfair.” Bias also includes “unfair.” You can see how fairness and bias are often confused. Machine learning systems can be affected at every stage. Let’s take an example: If you are building a machine learning model that predicts the next word from a series of texts, historical bias can affect your data. You give every book published in the past 50 year to ensure you have lots of data. Then you ask the model to predict this next sentence: “The CEOs Name is ______.” Unsurprisingly, you notice that the model predicts male CEO names more often than it predicts female. It is possible that you have unintentionally borrowed historical stereotypes from our society and incorporated them into your model. Bias can also be present in the model. Evaluation bias is when the model’s data doesn’t match the reality. This could be illustrated by training facial recognition software and testing it with photos taken on Instagram. Although your model may be very accurate on the test set, it will likely underperform in real life because most Instagram users are between 18 or 35.. The bias in your model towards this age group will make it perform less well on faces of younger or older people. Machine learning has many biases. I will cover each one in separate blogs. Although bias is often associated with negativity, it can also be used in machine learning. Knowing the problem before you start modeling can make it easier to choose relevant features. Although this can introduce bias, it may speed up or enhance the modeling process. Emily Morter, Unsplash Explainability. Also known as interpretability or explainability, this term describes how machine learning models make predictions. This is the process of interrogating models, collecting information about why certain predictions (or series) were made and then explaining this to people in a understandable manner. When explaining how a model functions, there are two scenarios you will encounter. You can only use the inputs and outputs of the model to create an explanation. White Box — Access to the model’s underlying data makes it easier to explain why certain predictions were made. White box models are generally simpler, and sometimes intentionally so explanations can easily be generated. However, a more simple and interpretable model may not capture all the data relationships. This could lead to a compromise between model interpretability or model performance. Explainability is usually based on one or two of the following: Model View — What features in general are most important to your model’s performance? What factors were responsible for a specific prediction? Explainability depends on your model’s black or white boxes, whether the instance or model views are relevant, as well as the data type you are exploring. These techniques are explained in detail by the Alibi open-source library. My personal preference is to refer to white-box model as “Interpretability”, because we need an interpretable model, and black-box as “Explainability”, since we attempt to explain the unknowable. Unfortunately, the official definition is not available and these terms are frequently used interchangeably. Piret Ilver, Unsplash Fairness Fairness Fairness by far is the subjectiveest of all three terms. Let’s look at bias as we did with bias. Now let’s see how machine learning applies it. This is often where AI can lead to disastrous outcomes. Fairness is crucial. A healthcare diagnosis or mortgage approval can have a profound impact on your life. It’s important to be fair with predictions. It’s likely you’re asking “What is a sensitive characteristic?”. This is an excellent question. It all depends on the type of sensitive you are. Examples of sensitive attributes include race, gender and sexual orientation. Although this may seem like an obvious thing to do, there could be multiple problems with it. The model might not function properly if the sensitive attributes are removed. Assume you are trying to predict how tall a child will become when they reach adulthood. Your predictions will be useless if you remove sensitive characteristics like age or sex. Fairness does not mean being an agnostic. It’s sometimes important to consider sensitive aspects in order to help people who may be marginalized by other characteristics. This is an example: university admissions. Raw grades may not always be enough to attract the most talented students. Higher scores might have been achieved by those who were able to access fewer education resources and had a better quality education. Other attributes might conceal sensitive features. Sometimes, it is possible to combine non-sensitive attributes with sensitive ones in order to find the value for sensitive features. A machine learning model might be able to determine the applicant’s race, gender, and nationality by simply knowing their full name. AI fairness can be a difficult topic. This requires that policymakers define fairness for every use case, which may sometimes prove subjective. There is often a compromise between individual fairness and group fairness. As we saw in the earlier example of university admissions, fairness for those with less privilege (group fairness), comes at the expense of others who have a strong educational background but whose grades are not good enough (individual unfairness). Summary: Fairness, explanationability and bias are not all the same.
