Need to understand the Ethical Bias of Pega Marketing 8.4
One of our client wants to upgrade from 7.x to 8.4. We are trying to demonstrate a new feature called Ethical Bias in Pega Marketing 8.4.
How does it actually work? I have configured ethical bias and able to run a simulation on top of it? Can you please illustrate with a basic example so that I build one use case around it.
Ex: I have 6 customers A, B, C, D, E, F. First 3 customers are male(A,B,C) and the other customers are female (D, E, F). Whenever I run an ethical bias simulation, it is always running successfully without any bias notification.
Is there any way to build some usecase based in the scenario? Is it expected to work like this?
***Edited by Moderator Marissa to update Platform Capability tags****
Thanks for your response. It helped me to understand the Ethical bias simulation. Now after running the bias simulation it was confirmed that there is a bias. There are two scenarios now.
Scenario 1 (Direct bias): I have a bias field called Gender which is used in the adaptive model. After running the simulation, I found that there is a bias shift on this. How do I fix this issue? Is this feature only limited to generate a simulation report? or Do we have any way to fix this one in business strategy or adaptive models?
Scenario 2 (Indirect bias): I have another field called Age which is not used in the adaptive model. Bias simulation predicts there is a bias shift based on the report. As it was not used in my adaptive model how do i fix this issues?
Do we have any ways to fix the above issues? Please let me know. Thanks in advance.
Posted: 2 years ago
Posted: 20 Apr 2020 11:27 EDT
Ivar Siccama (Ivar_Siccama)
Senior Product Manager, Machine Learning & AI
Indeed, the feature is currently limited to detecting any unwanted bias above threshold, using a simulation run.
For scenario 1, to remove the bias one would remove Gender from the list of potential predictors. At this point, there's no correction built in that would compensate for the gender bias (at the loss of predictive performance).
For scenario 2, it would be recommended to analyze which of the other predictors may be strongly correlated with Age. We have no built in direct capability yet that assists in this, although you may infer it from the predictor groups in an adaptive model instance that use correlations between predictors or use the built in machine learning when creating a predictive model which also detects correlations,