Brad Klingenberg, Director of Styling Algorithms, Stitch Fix
summary
Stitch Fix runs a two-stage system where models predict matches from client attributes, item attributes and feedback and human stylists pick the final five; this reduces wrong shipments but creates measurement headaches. Key claims: combining algorithms and curation improves outcomes but introduces selection bias that can censor failure modes, creates multiple success metrics (client buy rate versus stylist agreement), and causes non-compliance in experiments when humans override A/B rules, so careful logging, presentation-modeling and interpretable signals are required for it to work.
AI-written summary · about
chapters (5)
- 0:00Introduction to Stitch Fix
- 1:38The human-in-the-loop process
- 6:07Measuring success
- 11:40Defining the prediction goal
- 16:59Experimental complexity
related talks
25mPersonalization and Retail: Lessons from Stitch Fix
34mAugmenting Human Decision Making with Data Science
02mDBTB IN Eric Colson r
32mScaling Personalization: DevOps at Stitch Fix
42mAugmenting Human Decision Making with Data Science
42mAn Algorithm of Style
18mSynthesizing human and machine capabilities
40mEnabling simple ML ops at Stitch Fix