Improving Deep Learning for Airbnb Search (KDD 2020)
- Paper: Improving Deep Learning for Airbnb Search
- Venue: KDD 2020 (Applied Data Science track)
- Link: https://dl.acm.org/doi/abs/10.1145/3394486.3403333
- The paper documents Airbnb's experience bringing deep neural networks into a production search ranking stack that had long been dominated by GBDT models, an applied-industry case study rather than a new algorithm.
- It highlights the engineering friction between GBDTs and neural networks: differences in how features are consumed (raw values vs. discretization/normalization), and the need to rebuild feature pipelines to suit neural training.
- A key modeling contribution is a real-time user embedding that encodes the guest's in-session actions (listings clicked, stayed on, contacted), enabling the ranker to adapt to the user's current, evolving travel intent rather than relying only on historical long-term behavior.
- The team used the embedding within the existing ranking infrastructure via an incremental rollout strategy, mitigating the risk of a full model replacement.
- The paper stresses that offline ranking metrics do not fully predict online performance; Airbnb relied on online A/B tests and business metrics such as booking value to validate changes.
- It discusses negative-sampling challenges for training embeddings on sparse booking events, a recurring theme in recommendation and marketplace retrieval.
- An interpretable ensemble of graph and language models for improving search relevance
- Applying Deep Learning to Ads Conversion Prediction in Last Mile Delivery
- Behavior Modeling Space Reconstruction for E-Commerce Search
- Behavior-driven query similarity prediction based on pre-trained language models
Key points
Why it matters
This is one of the frequently cited industrial references on the practicalities of deep-learning adoption in large-scale search and recommendation: incremental deployment, feature-pipeline refactoring, real-time session representations, and the offline-online metric gap.