Evan Estola, Lead Machine Learning Engineer, Meetup, NYC 2017
summary
Meetup recommendation engineering faces extreme cold start sparsity from time and place so simple collaborative filters reproduce historical biases — women shown craft and skincare when creating business groups. The change is concrete: flip the model to predict topics for groups from group topics instead of predicting user interests, remove name/zip proxies, and add a long‑term join‑value objective. A fairness test tool detects accidental proxies; revenue objectives (Netflix style) and tradeoffs are measured rather than proclaimed.
AI-written summary · about
chapters (21)
- 0:00Intro
- 0:36Meetup
- 2:25Machine Learning Algorithms
- 2:58Gender Bias
- 3:34When Recommendations Go Bad
- 4:11Google Vision
- 5:15Microsoft Chat
- 5:32Racism
- 6:15Data Ethics
- 6:58Summary
- 7:09Statistics
- 7:19Bias
- 8:12Whats right
- 9:16Not all features are useful
- 9:59Offline performance is not guaranteed
- 10:55Pablo Picasso quote
- 11:10Asking the right questions
- 13:27Fair test algorithm
- 15:37questionable realworld applications
- 19:34the point
- 20:30old wisdom
related talks
47mAlgorithmic bias is everywhere (especially at Breeze) - what can we do about it?
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
47mEthical Machine Learning
36mMachine Learning Fairness: Lessons Learned
01h 30mFairness Tales: How To Measure And Mitigate Unfair Bias in Machine Learning Models
55mEthical Machine Learning
10mWhy diversity matters in Machine Learning communities
50mEthical Machine Learning