DataRobot Enterprise AI Platform: Bias and Fairness Testing

Showing results for 
Search instead for 
Did you mean: 

We're looking into an issue with broken attachments right now. Please stay tuned!

DataRobot Enterprise AI Platform: Bias and Fairness Testing

Automatic bias and fairness testing in the DataRobot Enterprise AI Platform ensures predictive models align with your ethics and values. It exposes and eliminates bias in your models, and prevents it from recurring.

This demo presents a real-world use case that illustrates the power of automated bias and fairness testing. Using DataRobot, we want to build a model that Megacorp (our hypothetical manufacturing company) will use for hiring. Importantly, the model needs to result in fair and unbiased predictions so that Megacorp truly hires the best candidates, regardless of age or gender.

Feature highlights

  • Identify and flag dataset features that need to be protected. For example, when building a model that helps predict the best candidates for historically male-dominated positions, you would flag the feature ‘gender’ so that it doesn’t sway the model’s results.



  • Select the fairness metric that makes sense to your use case and data. And if you don’t know which to choose, a short series of questions will help guide you, 


    and then recommend the metric to use.



  • Explore Per-Class Bias automated insights to understand if the model exhibits bias for your protected features.



  • Then, review Cross-class Data Disparity to understand which features are driving the bias, 

    and which of those features are causing the bias, so that you understand exactly how to mitigate bias from reoccurring.



What have we learned?

Notice the 40 and Over group shows two high blue bars indicating a large portion of the group has had less than 2 internships, and even no internships. Also see the four higher orange bars for the Under 40 group indicating most members of this group have had between 1 and ~5 internships. Clearly the number of ‘Internships’ exhibits a high amount of data disparity. This makes sense because applicants earlier in their careers are more likely to highlight internships versus those later in their careers (40 and Over applicants) who are more likely to highlight actual work experience.

If you have any questions or comments about what I’ve shown here, click Comment and let me know! I’m eager to hear your feedback.

More information

Version history
Last update:
‎02-02-2021 10:43 AM
Updated by: