Improving product discovery via relevance and ranking optimization
summary
Describes Flipkart’s two-layer recommendation pipeline: a relevance layer that computes product-pair similarity from user browsing and purchase events via collaborative filtering to build candidate sets, and a ranking layer that applies Learning-to-Rank models (with incrementally added engineered features) to predict conversion probability and produce the final ordered recommendations. Covers the team’s transition from rule-based approaches to a machine-learned system and the feature sets and algorithms used for ranking.
AI-written summary · about
chapters (13)
- 0:00Start
- 1:24The Curious Case of One Indian Girl
- 2:38The Recommendation Problem
- 4:28Relevance
- 7:44Pattern Mining: Computing Score between Products W
- 8:43Hierarchical Aggregation & Latent Concepts
- 11:05Attributes Similarity
- 13:13Visual Embeddings
- 18:45Ranking: Insights
- 20:16Big Billion Days!
- 22:17Product Quality Features
- 24:40Historical Features
- 30:23References
related talks
43mRecommendation System beyond traditional Collaborative filtering
21mAnalytics & Science for Customer Experience in E-commerce
21mAmazon Search: The Joy of Ranking Products
06mINTERVIEW
32mLeveraging recommender systems to personalise search results
38mRecommendations at Rappi: from MVP to Data Product
31mEverything in its Right Place: Optimising Ranking in Online Grocery
41mRecommender Systems: Crafting Effective Personalised Experiences