KOEN
Read Original
#AI

Recommending travel destinations to help users explore

Recommending travel destinations to help users explore
01

Summary

"Where to Next?" How Airbnb Recommends the Perfect Destination

Building a Smart Destination Recommendation Engine with Transformers and Geolocation Insights

This article unveils Airbnb's destination recommendation framework designed to convert vague exploration into concrete bookings. It shares technical insights into modeling long-term and short-term user interests as sequences and leveraging geographical hierarchies to suggest the best travel spots.

  • 01Modeling search, view, and booking history as sequence tokens using Transformer architecture
  • 02Enhancing geolocation understanding via multi-task learning for city and region-level predictions
  • 03Strategic training data design to balance behaviors between active and dormant users
  • 04Proven business impact through implementation in search autosuggest and re-engagement emails

RECOMMENDATION

A must-read for engineers building personalization systems based on user behavior sequences or those looking to integrate hierarchical data into multi-task learning models.

The Problem

Airbnb users in the early trip planning stage often lack clear destinations or dates, making it challenging to identify their ambiguous intent and reduce decision friction for booking.

The Solution

The team developed a Transformer-based model that treats user history as sequence tokens and introduced a multi-task learning framework to predict both city and region levels while balancing data for active and dormant users.

The Result

Deployment in autosuggest and abandoned search emails led to measurable booking gains, particularly in non-English regions, by helping users discover relevant and affordable destination alternatives.

Trade-off

The model's effectiveness is limited by user opt-outs from personalization, and the current scope primarily focuses on destinations, leaving timing and price integration for future iterations.

03

Key Concepts

Concept · 01

Transformer

A deep learning architecture that learns relationships in sequential data using attention mechanisms.

  • Used to model sequences of user actions (bookings, views, searches) to capture journey context.
Concept · 02

Multi-task Learning

A learning paradigm where multiple related tasks are learned simultaneously to improve generalization.

  • Implemented with multiple prediction heads for city and region levels to learn richer geolocation representations.
Concept · 03

Active vs. Dormant Users

A strategy to differentiate modeling approaches for users with recent activity versus those who haven't visited in a long time.

  • Training data was sampled differently for each group to mimic both late-stage booking and early-stage planning scenarios.