1
Warning: I spent 2 years optimizing a hiring algorithm for bias before realizing my training data was garbage
I was so focused on tweaking the weights and thresholds that I never checked if the historical hire data I was using actually represented good hires. A recruiter finally pointed out that 60% of our 'successful' hires from that data were just people who stayed the longest, not performed the best. Has anyone else found a huge flaw in their training set way later than they should have?
3 comments
Log in to join the discussion
Log In3 Comments
evahenderson1mo ago
Wait, was this recruiter just sitting on that info the whole time watching you tune hyperparameters like a reality show? Its almost worse when someone lets you go 2 years down a wrong path before casually mentioning the foundation is made of sand. I swear the real bias in machine learning is assuming your data was collected by anyone with common sense. At least now you can rebrand your algorithm as a "tenure predictor" and sell it to HR departments who confuse loyalty with performance.
3
Did your recruiter actually sit down and look you in the eye before dropping that bomb? I had a similar wake-up call last year when I realized my model for predicting customer churn was basically just flagging people who had already sent in a cancellation request. It's painful how easy it is to miss the obvious when you're deep in the weeds.
1