Our experiments with food recommendations
summary
Swiggy describes running production experiments that move from restaurant‑first to an item‑first dish discovery (launched in Bangalore) using a hybrid of matrix factorization, topic models over order histories, image classification and item/restaurant embeddings, plus pairwise learning‑to‑rank. Models incorporate real‑time serviceability/stress signals and demand‑shaping constraints and optimize relevance subject to availability. Results feed a live system where about 60% of orders come from feed and 30% from search, and the deployed hybrid improves relevance while reducing overload on hot restaurants.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:43Show of hands
- 2:05What is recommendation system
- 6:04What makes recommendation system unique
- 8:41Other factors to consider
- 11:57Data
- 14:41Problem
- 15:39Solution
- 16:42Item Level
- 19:17Questions
related talks
35mData Science for Dinner: Recommending Recipes
37mFood is personal! Design and impact of personalized recommendations in e-grocery
41mHow Deliveroo improved the ranking of restaurants
45mContent recommendation at iFood scale
20mJeremy Schiff, Senior Manager, Data Science, OpenTable
41mCapacity and efficiency problems in real-time logistics - and how data at Swiggy solves this
22mData-Driven F&B Delivery: Jahez as a Leading Example
39mDiscover your latent food graph with this 1...