An Algorithm of Style
summary
Thread scales eight stylists to about half a million clients by combining stylist-curated candidates with ML: a simple non-negative matrix factorization on users’ image picks produces human-interpretable style components that work well, while off‑the‑shelf collaborative filtering fails on sparse, churny inventory. A hybrid classifier predicts clicks/orders using user/item metadata plus short‑term event features. The production stack includes Spark streaming, Kinesis/S3 event storage and a real-time feature pipeline with cached aggregates in Redis and shared Python feature code.
AI-written summary · about
chapters (17)
- 0:00Thread and personal styling
- 3:18Talk overview and motivation
- 4:02Defining user style visually
- 7:37Non-negative matrix factorization for style
- 9:44Style component analysis benefits
- 13:58Collaborative filtering struggles
- 18:39Hybrid recommender model and real-time
- 21:30Real-time training pipeline challenges
- 26:11Real-time feature pipeline wishlist
- 31:15Shared code pipeline engineering
- 34:04Training data bias in order models
- 36:00Q&A: Stylists and ML models
- 37:10Q&A: Measuring recommendations
- 38:37Q&A: Women's service expansion
- 39:00Q&A: Seasonality in user taste
- 40:25Q&A: Preventing data leakage
- 41:29Closing remarks
related talks
35mSpotting trends and tailoring recommendations: PySpark on Big Data in fashion
18mSynthesizing human and machine capabilities
20mBrad Klingenberg, Director of Styling Algorithms, Stitch Fix
34mAugmenting Human Decision Making with Data Science
42mAugmenting Human Decision Making with Data Science
39mAnd Then There Are Algorithms
32mScaling Personalization: DevOps at Stitch Fix
28mThe Power of Personalized Customer Experiences Through AI Innovations