Specific Methods for Eliminating Bias in Machine Learning

Just like humans, data can be biased. In fact, bias in machine learning largely happens because machines will reflect human input. Machine learning (ML) has incredible capabilities and potential, but recognizing, preparing for, and adjusting for inherent bias is key to maximizing success and minimizing negative impacts.

Built In SF spoke to three companies, including Evidation Health, about how their engineers consider and prepare for bias in and the cons of ML.

The Research, Analysis, and Learning (ReAL) Team at Evidation says that when choosing the right learning model for a given problem, one of the most important things is to not only recognize that there is both human and machine bias—but also understand what the downstream effects of those biases will be. The team notes that bias can occur at any stage of the ML pipeline, including design, pre-processing, and modeling. And while it’s not always easy to detect, they use a variety of approaches to mitigate as much bias as they can, while recognizing that it may be impossible to eliminate all bias.

When it comes to making sure their training data set is diverse and representative, the ReAL team works to ensure proper epidemiological and econometric analytic handling of protected features like race and gender. “We do this across analyses by addressing confounding, mediation and effect moderation or modification whenever possible while implementing counterfactual fairness metrics in predictive modeling.”

A culture that encourages and empowers anyone to raise concerns is one of the things that the Evidation team says helps them identify and eliminate bias when it occurs. Team members are trained in bias, data ethics, and casual interference, which helps them understand how seemingly benign decisions can have real societal impacts. A diversity of backgrounds, experience, and ideas on the team is another key way in which Evidation constantly seeks to grow in a way that allows biases to be revealed and eliminated.

To read more about how Evidation Health, Grammarly, and Sift address bias in machine learning, read the full article on Built In SF.