This article was co-written with Sean Mealin.
“Unconscious bias is present in everything we do”
It’s a sentiment shared by psychotherapists and statisticians alike. Bias, conscious or not, permeates every one of our day-to-day thoughts and functions in some way or another. Though we like to think of AI as a cold and calculated machine, it can harbor some pretty significant biases if trained the wrong way. Biases bring massive consequences when it comes to the design of AI and Machine learning models and that can have significant downstream effects if left unaddressed. We’re glad to announce that SAS is taking another step in the fight for ethical AI by integrating automatic bias mitigation several of its most popular machine learning procedures.
What is bias?
Bias in machine learning refers to systematic errors in model predictions that result from incorrect assumptions, flawed data, or skewed algorithmic design. These errors can lead to unfair, inaccurate, or discriminatory outcomes.
It can be categorized into several types:
- Prediction Bias: When the average prediction of a model deviates from the actual ground-truth values
- Training Data Bias: Occurs when the data used to train the model is not representative of the real-world population (e.g., underrepresentation of minorities)
- Algorithmic Bias: Arises from the design of the model itself, such as over-regularization or optimization for accuracy at the expense of fairness
- Intersectional Bias: Discrimination against groups with multiple marginalized identities (e.g., Black women), which single-attribute fairness interventions often fail to address
The real consequences of bias in machine learning
In 2014, a Fortune 100 company was blasted in the media after training it’s AI model for recruiting using a decade’s worth of resumes from mostly male employees. As a result, their AI started favoring male candidates and penalizing resumes with words like “women’s” and downgraded graduates from women’s colleges meaning it became much harder for women to get roles at amazon than their similarly qualified male counterparts.
Another example of the real consequences of bias is a major health insurance provider currently navigating a class action lawsuit following allegations that their use of biased AI in determining insurance claim acceptance or denial. The lawsuit claims the company’s algorithm is wrongly denying major medical claims and leaving people to pay their medical bills out of pocket often landing them in significant financial strife.
Cutting bias out at the root
We at SAS are dedicated to delivering trustworthy AI and committed to removing bias from the equation. With the latest update to the software, SAS Viya now has bias detection and mitigation built into every machine learning procedure. This means less manual effort, more reliable models, and the comfort that you can trust you’re AI is making ethical decisions.
For bias mitigation, there are three main categories:
- Preprocess methods: tries to mitigate bias by altering the training data set before model training begins
- In-process methods: tries to mitigate bias by altering the model parameters during the training process
- Postprocess methods: tries to mitigate bias by altering the outputs of the model during scoring
When the MITIGATEBIAS option is used on a supported PROC, we run the exponentiated gradient reduction (EGR) algorithm. This algorithm is an in-process method. It works by adjusting weights of observations during the model training process.. This helps to reduce the model’s bias but also increases the model training time.
Delivering you trustworthy AI
Trust is at the core of what we deliver and SAS Viya helps you build trustworthy AI models responsibly and ethically. We are actively working to build better tools and programs to make it easy for any user to ensure their AI and Machine learning models are delivering reliable and unbiased outputs out of the box.
Looking for an easy way to figure out what our procedures do?
Download our Procedures Cheat Sheet
