The Necessity of Algorithmic Biases

Most of my articles thus far have focused on the negative effects of biases on social media. When I talk about biases on social media and in social media algorithms, I am referring to user’s feeds, which are controlled by recommendation algorithms, only showing content that reflects the user’s ways of thinking. An example of how this can be harmful would be a user who might have interacted with some content that stereotypes marginalized races and is then only recommended more racist content, furthering their racist beliefs. Other ways in which biases in recommendation algorithms could cause harm would be through promoting political polarization, stereotypical or hateful content, or a narrow world view. Even when not promoting hateful ideologies, a feed that only recommends content that aligns with the user’s views can get boring, and is not conducive to the formation of empathy and understanding between people that are different. When we talk about biases, especially in humans, normally we are talking about negative ones like I just mentioned. However, in machine learning, biases are unavoidable, and not always harmful.

In a conversation with Harvey Mudd College Professor of computer science, George D. Montañez, we discussed how biases are integrated into machine learning. Dr. Montañez’s specialty is theoretical machine learning, he specifically focuses on the limits of machine learning and how assumptions control how machine learning works. Biases, in terms of machine learning, are essentially assumptions made by the algorithm. Any learning algorithm is going to learn from given data and prepare for future data it will encounter, through making predictions and therefore assumptions. Dr. Montañez also explained the difference between responsible and irresponsible biases in machine learning. An irresponsible use of assumptions in ML algorithms would be a lack of care applied to what assumptions one is building into an algorithm. For example, training facial recognition technology with only caucasian faces. Whereas responsible assumptions allow the algorithm to make accurate predictions beyond the training set. Following up on the last example, responsible training of the facial recognition algorithm would include a diverse array of faces in the training set. Dr. Montañez provided the example of popular online chat bots, which can fall into the pitfalls of uncritically taking in data sources from the internet.

In the context of recommendation systems, these algorithms aim to make assumptions about what kind of content a user enjoys. Generally, recommendation algorithms assume that similar users will like similar content. These assumptions are based on a variety of other factors as well, such as, gender, age, location, and what content the user interacts with. The algorithm then performs complex math with the values or factors, it is provided with. It’s also possible that there are hidden values involved in this math that we are unaware of. One risk associated with recommendation engines is that the algorithms will make stereotypical assumptions, like assuming what someone will like based on their race.

The word “bias” has bad connotations when applied to humans. When a human is biased, they hold prejudice against something or someone, they prefer their own over others. However, in machine learning, bias is more of a necessary function. Machine learning algorithms must have some form of bias in order to make assumptions about data beyond the training set. These biases can be implemented in a responsible and irresponsible way, however they will always be present. In social media algorithms, assumptions produce our curated explore feeds, that usually are quite enjoyable. They also have the ability to curate these feeds in a stereotyped way. Creating machine learning algorithms requires attention to detail, and purposeful training of the algorithms.