How Airbnb Uses a Transformer Sequence Model to Personalize Search by Learning Guest Journeys

Challenges

Guest event streams are dominated by browsing actions; some guests accumulate tens of thousands of events, making raw sequence modeling computationally infeasible. Booking events are sparse but high‑value, requiring the model to separate genuine intent from noisy clicks. Training on billions of search‑tag pairs incurs high cost, demanding efficiency optimizations.

Solution Architecture

The timeline for each guest is split into two sequences:

Long‑term sequence: up to 80 events covering the past seven years, capturing low‑frequency but informative actions such as bookings, reviews, and cancellations.

Short‑term sequence: up to 200 events from the past 21 days, capturing recent browsing behavior.

These limits truncate only the longest 2 % of sequences. Both sequences share a unified embedding table for high‑cardinality identifiers (property ID, host ID, hierarchical geographic IDs) as described in [1] .

Training and Inference Efficiency

Three strategies raise encoder throughput by roughly fourfold:

Search batching : the encoder uses causal masking, so a single forward pass over the full event timeline produces embeddings at every timestamp. For a guest issuing searches at T=4, T=7, and T=10, the embeddings generated at T=3, T=6, and T=9 can be reused for the three searches, avoiding separate encoder runs.

Length‑based bucketing reduces zero‑padding within batches.

A sparse‑computation mode removes padded searches from the ranking model.

During inference, the encoder runs offline as a daily batch job, storing per‑guest embeddings. At query time the ranking model retrieves the stored embedding, concatenates it with the current query, and scores candidate listings, keeping latency low while leveraging the full guest history.

Rollout Milestones and Evaluation

Three rollout phases were evaluated with rigorous A/B testing:

Phase 1: only the long‑term sequence.

Phase 2: add the short‑term browsing sequence, applying the efficiency tricks above.

Phase 3: introduce a set‑wise ranker trained jointly with the sequence encoder ( [2] , [3] ), allowing simultaneous comparison of candidate listings.

Results

Offline experiments:

Long‑term sequence alone: +0.44 % NDCG.

Both sequences: +1.48 % NDCG.

Set‑wise ranker: additional +2.3 % NDCG, total +3.78 %.

Online A/B tests (statistically significant):

Long‑term only: completed bookings +0.31 %, views +0.38 %.

+ Short‑term: completed bookings +0.55 %, nights booked +0.82 %, views +0.90 %.

+ Set‑wise ranker: completed bookings +0.28 %, booking initiations +0.32 %.

Applying the same guest embeddings to promotional‑email ranking (no architectural changes) yielded +5.04 % email click‑through rate, +0.16 % completed bookings, and +0.23 % nights booked.

Conclusion and Future Work

Encoding the full guest journey with a Transformer sequence model provides richer guest representations and substantially improves search relevance and conversion. Ongoing work explores near‑real‑time embedding updates, richer event types (wish‑lists, map interactions), intent‑aware ranking, generative recommenders, and further refinements to the set‑wise ranker. For full technical details see the KDD 2026 paper “JourneyFormer: Encoding Airbnb Guest Journey with Sequence Modeling” [4] .

Code example

[1] Coleman, Benjamin, et al. “Unified embedding: Battle-tested feature representations for web-scale ML systems.” Advances in Neural Information Processing Systems, 2023.
[2] Tang, Jie, et al. “Learning to Comparison-Shop.” ACM International Conference on Information and Knowledge Management. 2025.
[3] Haldar, Malay, et al. “Beyond Pairwise Learning-To-Rank At Airbnb.” ACM International Conference on Information and Knowledge Management. 2025.
[4] Zha, Daochen, et al. “JourneyFormer: Encoding Airbnb Guest Journey with Sequence Modeling.” ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2026.