Problem
Myntra, an Indian fashion e-commerce platform, lists millions of items. Even after category and attribute filters, a user still faces thousands. Recommendation models trained on a user's own history cannot serve a first-time visitor, which is the user cold start problem. Fashion also differs by region: climate, terrain and local clothing culture change what people buy. A single national ranking misses this, and the team needed a location signal that works for new and existing users alike.
Example from the article: in winter, the "Trending near you" men's wear widget shows winter wear in Uttar Pradesh and light everyday wear in Telangana.
Solution in brief
Treat the delivery pincode as the "user". Build a pincode-by-item matrix of "at least K orders", train Bayesian Personalized Ranking (BPR) on it to get pincode and item embeddings, and cluster the pincodes. The embeddings and clusters then feed ranking features and a regional trending widget.
Questions this case answers
- How to give a user with no history a personalized item list when the only known attribute is a coarse category of where they are
- How to learn a reusable representation of groups of users that share a context, when per-user data is too sparse or missing
- How to recommend relevant products to a new user who has no interaction history?
- How to show products that are trending in the user's own region?
Tools and methods
- Bayesian Personalized Ranking (BPR) — pairwise collaborative filtering that learns pincode and item embeddings
- Pincode-by-item interaction matrix — binary input: 1 if the item has at least K orders from the pincode
- Clustering of pincode embeddings — groups similar pincodes, used for cluster-level popularity
- nDCG@k per pincode cluster — offline check that cluster popularity beats national popularity
- Negative sampling — random, or weighted by item popularity
Google or a one-time email code. No password.
