From Data to Insights: Segmenting Airbnb's Supply

Alexandre Salama, Tim Abraham25 Nov 2024

Problem

Airbnb hosts differ in earnings goals and schedule constraints, so listings are open for booking in very different ways. Products, campaigns and operations must treat a seasonal host and a full-time host differently. A single feature misleads: about 30% availability can mean weekends only or one summer season. Classic segmentation such as RFM (Recency, Frequency, Monetary) targets customer value, not calendar dynamics, and is usually a one-off analysis on a small dataset. Airbnb needed segments for millions of listings, refreshed daily.

Example from the article: Alice lists a 2-bedroom apartment only in summer, Max keeps an identical one open all year. Alice gets seasonal pricing tips, Max gets booking analytics and tools for many reservations.

Solution in brief

Build four calendar features per listing and cluster a large sample with k-means into 8 named segments. Then train a decision tree on the cluster labels and translate it into a SQL CASE WHEN query that scores all listings in the warehouse.

Questions this case answers
  • How to segment entities by temporal activity patterns when a single summary rate hides different calendar shapes and no ground-truth labels exist
  • How to deploy a clustering model for daily scoring of millions of records inside a SQL warehouse
  • How to segment supply-side users by calendar availability patterns at the scale of millions of daily records?
  • How to tell apart two hosts with the same availability rate but different calendars?

Tools and methods

  • Availability rate and streakiness features — describe volume and consistency of a calendar
  • Seasonality features — quarters with availability and longest run of available months
  • k-means with an elbow plot — find 8 clusters in the scaled features
  • Decision tree distilled from cluster labels — interpretable rules that become SQL
  • A/B test, correlates and UX research — validate segments without labels
Continue reading — it's free

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