Research Paper

Core Innovation

This paper presents JourneyFormer, a sequence modeling approach tailored for complex and sparse guest booking sequences at Airbnb. It introduces design strategies for event selection, ID embeddings, model architecture, and label attribution to balance effectiveness and scalability in production. The model also incorporates training and inference acceleration techniques for real-world deployment.

Why It Matters

Accurately modeling user journeys in platforms like Airbnb is critical for delivering relevant search results and increasing bookings. JourneyFormer improves recommendation effectiveness despite sparse booking data and complex user behavior, leading to better user experience and higher revenue. Its scalable design supports deployment in large-scale production environments, benefiting online marketplaces broadly.

Potential Customers & Pain Points

• Online travel marketplaces – Difficulty modeling complex sparse user booking data
• E-commerce platforms – Need to improve recommendation relevance from long user behavior sequences
• Digital advertising networks – Require better user intent prediction from sparse signals

Business Model

SaaS or API-based platform offering sequence modeling and ranking optimization tools to online marketplaces and e-commerce companies, with pricing based on usage and model customization.

Competitive Landscape

• Google Recommendations AI
• Amazon Personalize
• Coveo
• Algolia

Implementation Challenges

• Handling extremely sparse and noisy user behavior data
• Balancing model complexity with production scalability
• Integrating sequence models into existing ranking pipelines

Validation Strategy

• Conduct offline evaluations comparing ranking metrics against baseline models
• Run online A/B tests measuring key business metrics such as booking rates and revenue
• Pilot deployments with select customers to gather feedback and optimize integration