Problem
Online fashion shoppers cannot try clothes on, so they hesitate over the size or pick the wrong one. A wrong size costs a sale or a return. Myntra, an Indian fashion retailer, needs a size suggestion on every product page. A model trained on purchase history covers only the item types a user bought before, and first-time users have no history at all. The service must also answer fast under peak-sale traffic.
Example from the article: a customer who bought a shirt within the last 2 years gets a shirt size from the purchase-based model, but that model cannot suggest a trouser size.
Solution in brief
Four models cover the cases: size from purchase history, size for categories never bought, size for first-time users from a questionnaire, and a fit percentage on top. Offline jobs precompute vectors, and an online service reads them and answers within 50 ms.
Questions this case answers
- How to predict the best option from a discrete ordered set for each user when options are not comparable across items and the user has history only for some item types
- How to personalize a prediction for users with no history using a short questionnaire and answers from related categories
- How to serve model-based recommendations within a tight latency budget and monitor the share of requests that get no answer
- How to recommend the right clothing size to a shopper from past purchases?
Tools and methods
- Personalized Size Recommendation (PSR) — scores each size of the viewed style from purchase history
- Meta Learning model — sizes for never-bought categories: attribute embeddings plus linear regression
- Cross-Learning model — first-time users: "Try Size Finder" answers reused across related categories
- Fit Score — fit percentage on top of the size output
- Offline pipelines and online service — refresh vectors, serve recommendations within a 50 ms SLA
- A/B tests and leakage dashboard — adherence, returns and coverage metrics
Google or a one-time email code. No password.
