Graph Machine Learning at Airbnb

Devin Soni14 Jun 2022

Problem

Trust and safety models usually score a user from features of that user alone: location, account age, number of bookings. For a new host these say little about trustworthiness. With no further signal the platform may push the host through a slower onboarding or ask for a lot of information up front. Models that treat users as independent rows ignore where the user sits among other users.

Example from the article: a new host shares a listing photo with a tenured host and lists in the same house as three other hosts. Two hops out, many hosts at that location share a business name. The new host is very likely part of an existing group renting rooms in one house.

Solution in brief

Build a user graph and turn each node's neighborhood into an embedding that mixes structure and neighbor features. Airbnb uses SIGN with weight-free aggregators (SGC), so the graph part needs no training. Embeddings are computed in a periodic offline job, stored in a feature store and passed to downstream models as ordinary features.

Questions this case answers
  • How to add relational context between entities to a prediction model when the entity's own attributes are too sparse to judge it
  • How to serve graph-derived features to a low-latency model without computing the graph embedding at request time
  • How to add information about a user's connections to trust and safety models when the user's own features are thin?
  • How to judge a new user's trustworthiness before the platform has any history on them?

Tools and methods

  • Graph convolutional network (GCN) — aggregates a node's neighborhood with its own features.
  • SGC and SIGN — fixed aggregations without trainable weights, precomputed and fed in parallel to a downstream model.
  • Offline batch pipeline and feature store — all embeddings computed at once; online models look them up.
  • SHAP — checks that embedding components matter in downstream models.
Continue reading — it's free

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