How to Remove Unfair Bias From Your AI – DataRobot Demo
So, you know about AI biases but want to see a demonstration of what’s involved in identifying and removing them from a machine learning (ML)/AI application? A recent webinar by DataRobot does just that: it walks you through a small ML project and explains step by step what to do and how. (The demo starts about 25 minutes into the recording.)
The project’s goal is to predict which job applicants should be hired by a fictitious high-tech firm, looking at their historical hiring data. Among DataRobot’s capabilities is a set of features that help you visualize and inspect various dimensions and components of the data and the predictive model built from that data.
For example, the graph below ranks the features used – internships, educational level, extra-curricular activities, gender (at the bottom of the screenshot), etc. – based on how much they weigh in predicting whom we should hire. (Note: Gender should not be used in ML as it is a sensitive attribute. It is used in the demo for illustration purposes.)
All images: How to remove unfair bias from your AI, webinar, DataRobot.
You can also zoom in to any of the features and see what is going on there. For example, a person is more likely to get hired if they had at least one internship:
You can further analyze which features predict highest at the individual level (the rows below the graph are individual records/job applicants):
When dealing with textual data – e.g. extra-curricular activities – you can see by way of a word cloud which words and phrases are proxies for gender. (Here, red indicates male and blue indicates female.)
Clearly, some sports names are highly correlated with gender, and so we can either remove this feature from the model completely or use it in a modified form: activity counts instead of names, how many different activities a “desired” employee has on average listed on their resume, and so on.
Our Take
While nothing in the above demonstration is earth shattering, we appreciate the clear, easy-to-use interface and the fact that these features are part of the DataRobot ML/AI platform. And we appreciate that they have been explained through the example, which is easy to understand and relate to, without much technical jargon.
This demo also sheds some light on the tasks and activities involved in machine learning and its iterative nature. (ML and data science are many times more art than science.)
All this means that you now have a better understanding of the approach and the mechanics of removing hidden biases and that you can start addressing them before they harm your customers or your organization.
Want to Know More?
Not all AL biases, however, come from data. To learn about other types of AI biases – those stemming from algorithms used, project design, team composition, etc. – and to learn about how identify and mitigate them, consult Info-Tech’s blueprint Mitigate Machine Bias.