Airbnb knows your taste after one session
How Airbnb turned 800 million search sessions into listing embeddings, and the trick that handles brand-new listings
Beat by beat.
| t (s) | Voice-over | On screen | Caption |
|---|---|---|---|
| 0–3 | You clicked four Airbnbs. Airbnb now knows your type. | Screen recording of 4 clicks; map dots light up | FOUR CLICKS |
| 3–7 | I read the paper Airbnb published on this. One trick handles brand-new listings nobody has clicked yet. Guess how. | Paper S3 header | NEW LISTINGS? |
| 7–14 | Search ranking and similar-listing recommendations, the two channels in the paper, drive 99% of conversions. [S3] | Number card | 99% |
| 14–24 | They trained on more than 800 million search click sessions. Each listing became 32 numbers, an embedding, so listings browsed together sit close together. [S3] | 3D scatter forming clusters | 800M SESSIONS |
| 24–32 | It learned what nobody typed: clusters matched location, type and price, and the authors checked for architecture, style and feel. [S3] | Clusters on a California map (recreated) | STYLE AND FEEL |
| 32–40 | For similar listings, the embeddings beat the old algorithm's click-through rate by 20%. [S3] | Bar 100 → 120 | +20% CTR |
| 40–48 | The trick: a new listing takes the average of the 3 nearest listings of the same type and price bucket. That covers more than 98% of new listings. [S3] | 3 dots average into a new one | BORROWED TASTE |
| 48–55 | Brew principle: cold start is a neighborhood problem. Lend new things the reputation of their honest neighbors. | Monochrome card | NEIGHBORS |
| 55–60 | Next: a sentence Amazon has used for over 20 years. You read it today. Guess it below. | Blurred “Customers who…” | NEXT: AMAZON |
On a phone, swipe the table sideways. Every factual line carries a source chip.
How it works.
Mechanism: listings are treated like words and click sessions like sentences; a skip-gram-style model learns a 32-dimensional embedding per listing from 800M+ sessions.
Evidence it captures taste: clusters line up with location, listing type and price, and the authors inspected whether architecture, style and feel come through.
Cold start: a new listing borrows the mean embedding of its 3 nearest same-type, same-price-bucket neighbours, covering 98%+ of new listings.
What a great PM would do: never launch a new item with a blank reputation. Lend it a fair, explainable prior and let real behaviour take over.
Steelman: The system matches guests to homes they would actually book: two-sided value, not just clicks.
Cold start is a neighborhood problem.
Sources
[S3] M. Grbovic & H. Cheng, “Real-time Personalization using Embeddings for Search Ranking at Airbnb”, KDD 2018 · https://doi.org/10.1145/3219819.3219885