SAS Viya Integrates Automatic Bias Mitigation in ML Procedures
Bias, whether conscious or unconscious, is an inherent part of human thought and function, and it poses a significant challenge when designing artificial intelligence and machine learning models. Far from being cold, impartial machines, AI systems can harbor profound biases if trained on flawed data or with skewed algorithms. Such biases carry substantial consequences, potentially leading to unfair, inaccurate, or discriminatory outcomes if left unaddressed. In a significant move towards fostering ethical AI, SAS has announced the integration of automatic bias mitigation into several of its most popular machine learning procedures within SAS Viya.
In machine learning, bias refers to systematic errors in model predictions stemming from incorrect assumptions, flawed data, or deficiencies in algorithmic design. These errors can manifest in several ways. Prediction bias occurs when a model’s average prediction deviates consistently from the actual ground-truth values. Training data bias arises when the dataset used to train the model does not accurately represent the real-world population, such as an underrepresentation of minority groups. Algorithmic bias, on the other hand, originates from the design of the model itself, perhaps through excessive regularization or an optimization strategy that prioritizes accuracy at the expense of fairness. A particularly insidious form is intersectional bias, which involves discrimination against groups with multiple marginalized identities—for instance, Black women—and is often missed by single-attribute fairness interventions.
The real-world ramifications of machine learning bias are already evident. In 2014, a major Fortune 100 company faced widespread media criticism after its AI recruiting model, trained on a decade’s worth of resumes predominantly from male employees, began favoring male candidates. This system penalized resumes containing words like “women’s” and downgraded graduates from women’s colleges, making it considerably harder for qualified women to secure roles compared to their male counterparts. More recently, a prominent health insurance provider is navigating a class-action lawsuit alleging that its AI-driven system for determining insurance claim acceptance or denial is biased. The lawsuit claims the algorithm is wrongly denying major medical claims, forcing individuals to pay out-of-pocket and often leading to significant financial distress.
Recognizing the critical need for trustworthy AI, SAS is committed to addressing bias at its root. The latest update to SAS Viya integrates bias detection and mitigation directly into its core machine learning procedures, aiming to reduce manual effort and enhance model reliability. This built-in capability offers users greater confidence that their AI models are making ethical decisions.
Bias mitigation strategies generally fall into three categories. Preprocess methods attempt to mitigate bias by altering the training dataset before model training commences. In-process methods, conversely, work by adjusting model parameters during the training process itself. Finally, postprocess methods aim to mitigate bias by altering the outputs of the model during the scoring phase. When the MITIGATEBIAS option is activated on a supported procedure within SAS Viya, the system employs the Exponentiated Gradient Reduction (EGR) algorithm. This is an in-process method that functions by adjusting the weights of individual data observations during the model training phase. While effective in reducing model bias, it is important to note that this approach can also increase the model’s training time.
SAS emphasizes its commitment to delivering trustworthy AI, with SAS Viya designed to help users build responsible and ethical AI models. The company states it is actively developing better tools and programs to ensure that AI and machine learning models consistently deliver reliable and unbiased outputs straight out of the box, affecting procedures such as FOREST, GRADBOOST, NNET, SVMACHINE, and TREESPLIT, among others.