Personalizing Airbnb search by learning from the guest journey

This post details the implementation of a Transformer-based sequence modeling system for Airbnb search personalization. It addresses challenges of modeling long and sparse event sequences by splitting them into long-term (bookings, reviews, cancellations) and short-term (listing views) components. Efficiency strategies for training include batching of searches, sequence length bucketing, and sparse calculation of searches. Serving efficiency is achieved by decoupling inference into a daily batch job for sequence encoding and real-time retrieval for ranking. The introduction of a setwise ranker, co-trained with the sequence encoder, further enhances personalization by considering candidate listings collectively.