Optimizing Airbnb Search Journey with Multi-task Learning (SIGKDD 2023)
This post indexes an applied research paper published at KDD 2023 by Airbnb, available via the ACM Digital Library:
- Paper: Optimizing Airbnb Search Journey with Multi-task Learning
- Venue: SIGKDD 2023
- Source: https://dl.acm.org/doi/abs/10.1145/3580305.3599881
- Type: Academic / industrial applied paper
- Journey-aware modeling: Search at Airbnb is framed as a multi-stage funnel; optimizing only per-session metrics misses downstream outcomes such as booking requests and completed trips.
- Multi-task learning: Multiple related tasks (e.g., predicting engagement vs. conversion) are trained with shared representations, letting early-funnel signals inform late-funnel predictions.
- Production setting: The work describes the architecture, training pipeline, and trade-offs involved in deploying MTL models in a real search stack serving millions of users.
- Evaluation: Results are validated through offline metrics and online A/B experiments, reflecting industrial best practice rather than benchmark-only evaluation.
- An interpretable ensemble of graph and language models for improving search
- 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 (2025)
- Behavior-driven query similarity prediction based on pre-trained language models
- Better to Ask in English: Cross-Lingual Evaluation of Large Language Models
- Original paper: https://dl.acm.org/doi/abs/10.1145/3580305.3599881
Overview
The paper addresses a core problem in large-scale marketplace search: optimizing not just individual search sessions, but the entire search journey of a guest — from initial queries and listing impressions to host contacts and eventual bookings. Traditional single-task rankers optimize immediate engagement signals, which may not correlate with long-term trip success. Airbnb's approach applies multi-task learning (MTL) to jointly model multiple stages and objectives of the search journey.
Key points
Why it matters for search/recommendation practitioners
1. Architecture: Multi-objective ranking with shared backbones remains a dominant pattern in industrial search and recommendation; this paper is a concrete marketplace case study. 2. Data: Session and booking logs encode the journey structure; label definition (what counts as "success") is a first-class design decision. 3. Evaluation: Offline ranking metrics alone are insufficient — online experiments measuring journey-level outcomes are essential. 4. Engineering: Latency budgets, training cost, and objective balancing (loss weighting across tasks) are the hard constraints when shipping such models.