Beyond the Click: Elevating DoorDash's personalized notification experience with GNN recommendation

Nimesh Sinha25 Jun 2024

Problem

DoorDash sends push, email and hub notifications that suggest restaurants a user has not tried. Many recipients are inactive, so they leave little interaction history. The signals that describe them are also spread over several sources: orders, views, user and restaurant attributes. A model trained on hand-built features from these tables misses relations between users and restaurants. The team needed better recommendations to raise notification engagement and monthly active users.

Example from the article: a "Try Something New" notification lists restaurants chosen for each user by a graph neural network (GNN).

Solution in brief

Users and restaurants become nodes of a heterogeneous graph, and orders and views become edges. An ID-aware GNN (ID-GNN) learns user and restaurant embeddings by predicting whether a user orders from a restaurant within seven days. A daily batch job takes the closest restaurants for the user's address, ranks them by embedding affinity and filters the result.

Questions this case answers
  • How to personalize a once-a-day outbound message for users with sparse history when the relevant signals are spread over several interaction tables
  • How to build entity embeddings from a heterogeneous interaction graph and use them to rank geographically constrained candidates in a batch pipeline
  • How to pick restaurants for push, email and hub notifications that bring inactive users back?
  • How to recommend new restaurants when a user has few interactions and the data sits in several tables?

Tools and methods

  • Heterogeneous graph — nodes are users and restaurants, edges are orders and views, built from relational tables
  • ID-GNN (Identity-aware GNN) — GNN variant that gives the center node its own message parameters; best offline model
  • Link prediction — training task: probability of a user ordering from a restaurant in the next 7 days
  • LightGBM — production baseline the GNN was compared against
  • Geohash candidate retrieval — up to N closest restaurants around the user's recent addresses
Continue reading — it's free

Google or a one-time email code. No password.