Does fairness in Machine Learning algorithms truly exist? When we talk about artificial intelligence, we mean the machine’s ability to perceive and respond with independence and act. This response happens in a way that would typically require human intelligence and the decision-making process environment. Machine learning is already being used in various ways in our lives. Every time we talk on our smartphones, looking for pictures or asking for restaurant suggestions, we interact with machine learning algorithms. They extract large amounts of raw data, such as the entire text of a book or the entire archive of a newspaper, and analyze patterns to extract information that may not be visible to human analysts.
But when these sets of big data include social bias, machines learn that too.
The challenge is to develop ML models whose accuracy and fairness could reach the optimum level. Let’s analyze an example of using a common machine learning algorithm to create the so-called “embedded words”. Each word is incorporated or corresponds to a point in “space”. Words that are semantically related are attributed to “points” that are close to each related “space”. This type of integration makes it easy for computer programs to recognize word relationships quickly and efficiently, but having this type of learning formulation creates some serious problems.
Asking the system to automatically generate large numbers of “It’s X as it’s Y” ratios, returns a lot of “clean data”. At the same time, it comes with answers that reflect clear stereotypes and biased “behavior”. This happens because it has related words. “he: the doctor, she: the nurse”, “he: the architect – she: the interior designer”. The fact that the machine learning system started as the equivalent of a newborn baby is the strength that allows him to learn interesting patterns. It is also its inability to fall victim to these blatant gender stereotypes. The algorithm makes its decisions based on words that often appear close to each other. If the given data/documents do reflect gender biases the model will learn these prejudices too. If they have the word “doctor” more often next to the word “he” than near “her” and the word “nurse” more often near “her” than “self”.
Using machine learning algorithms to reduce the “bias“
The algorithm can not only reflect the prejudices of society, but the system may potentially reinforce gender stereotypes. Fortunately, there are ways to use the machine learning algorithm itself to reduce its own bias. The debiasing system uses people to identify examples of the types of connections that are appropriate and those that need to be removed. Then, using these human-made discriminations, we quantified the extent to which gender was a factor in these words. We told our ML algorithm to remove the gender factor from the links in the integration. This removes biased stereotypes without reducing the overall usefulness of the integration. When this was done, the algorithm no longer showed gender stereotypes. This process can be applied to relevant ideas to eliminate other types of prejudices, such as racial or cultural stereotypes. We hope we got you closer to understanding if fairness in Machine Learning algorithms truly exists.