Learning to Rank for Maps at Airbnb (KDD 2024)
Source: https://dl.acm.org/doi/abs/10.1145/3637528.3671648 Venue: KDD 2024 | Category: Verticals / Search & Ranking
Overview
This paper describes Airbnb's work on applying Learning to Rank (LTR) to its map product. Unlike conventional list-based search, Airbnb's map experience requires ranking listings that are browsed spatially: the ranking model must consider not only the relevance of individual listings but also how results are presented within the visible map region a user is exploring.
Key points
- Problem setting: Ranking for a map-based browsing surface, where the result set and the map viewport interact — a scenario distinct from classic query-to-list ranking.
- Approach: Learning-to-rank techniques applied within Airbnb's production search stack, using user interaction signals (clicks, bookings) as training data.
- Industrial constraints: The system must meet live-traffic requirements for latency, scalability, and product-quality considerations beyond offline ranking metrics.
- Takeaway for practitioners: The paper serves as an industrial case study of adapting LTR to geospatial/vertical search, complementing the broader literature on retrieval, re-ranking, and recommender systems.
- An interpretable ensemble of graph and language models
- Applying Deep Learning to Ads Conversion Prediction in Last Mile Delivery
- Automated Query-Product Relevance Labeling using Large Language Models
- Behavior Modeling Space Reconstruction for E-Commerce Search
- Behavior-driven query similarity prediction based on pre-trained language models
> Note: The original abstract text was not fully available in the source material. For the complete method description, experiments, and quantitative results, please consult the official ACM publication (DOI: 10.1145/3637528.3671648).
Related entries in this collection
Glossary
| Term | Meaning | |------|---------| | LTR | Learning to Rank | | nDCG | Normalized Discounted Cumulative Gain, a ranking quality metric | | Ranking cascade | Pipeline of candidate retrieval followed by fine-grained re-ranking |
Recommended reading paths
1. Researchers: Compare the map-ranking formulation with standard LTR benchmarks and check how statistical significance and compute costs are reported. 2. Engineers: Identify reusable components (encoders, re-rankers) and evaluate integration cost with existing search stacks. 3. Product managers: Focus on user-perceivable benefits (latency, result relevance on the map) rather than offline metrics alone.