Problem
Zomato recommends restaurants and dishes to millions of customers. Classical recommenders build features from a customer's past behaviour, so they see only the restaurants and dishes that the customer touched. Most customers order from a few restaurants, so the data is sparse. The same restaurants and dishes also keep coming back, which is redundancy. The team wanted a richer signal for its personalisation model.
Example from the article: a customer who orders Mediterranean food may also like continental dishes, but a model fed only that customer's own history cannot see the link.
Solution in brief
Treat customers and restaurants as nodes of an interaction graph and learn node embeddings with GraphSAGE, unsupervised. Use the cosine similarity between a customer embedding and a restaurant embedding as one more feature in the existing personalisation model. It became the second most important feature by SHAP and slightly raised AUC, but training cost and per-city models made it impractical at scale.
Questions this case answers
- How to enrich a ranking model with similarity features learned from a sparse bipartite interaction graph when users touch only a few items
- How to decide whether a graph embedding model is worth its training time when it must be maintained separately for many regions
- How to add graph-based customer and restaurant embeddings to a personalization model?
- How to get recommendation signal for customers who interacted with only a few restaurants?
Tools and methods
- GraphSAGE — inductive node embeddings from sampled k-hop neighbourhoods
- Random-walk positive pairs and random negative pairs — unsupervised training signal
- Cosine similarity of embeddings — customer-restaurant feature for the downstream model
- SHAP — importance of the new feature in the personalisation model
- MRR, Precision@k, Recall@k — measures for the follow-up experiments
Google or a one-time email code. No password.