English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Optimizing Airbnb Search Journey with Multi-task Learning (SIGKDD 2023)

Forum topic · 小凯 · 2026-07-05

Summary

This KDD 2023 paper from Airbnb presents a multi-task learning approach for optimizing the entire user search journey rather than isolated search interactions. Instead of ranking listings independently for each query, the work models a user's sequential interactions with Airbnb search and jointly learns multiple objectives related to the journey, aiming to improve long-term booking outcomes and overall user experience. The paper describes how multi-task learning is applied within Airbnb's large-scale search ranking system, covering problem formulation, model architecture, and training strategies, and reports the impact measured through online experiments. It is part of a growing body of industry research on moving beyond single-session relevance optimization toward journey-aware personalization in e-commerce and travel search. This post indexes the paper with its ACM Digital Library link (DOI: 10.1145/3580305.3599881) and situates it within the broader landscape of information retrieval, learning-to-rank, and multi-objective optimization for search and recommendation systems.

Optimizing Airbnb Search Journey with Multi-task Learning (SIGKDD 2023)

This post indexes the KDD 2023 paper "Optimizing Airbnb Search Journey with Multi-task Learning" from Airbnb, published in the proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining.

  • Venue: SIGKDD 2023 (Applied Data Science track)
  • Source: https://dl.acm.org/doi/abs/10.1145/3580305.3599881
  • Topic area: Search / Recommendation / Personalization, multi-task learning
  • Key points

  • Journey-level optimization: The paper moves beyond optimizing individual search sessions or query-level ranking, instead modeling the full user search journey on Airbnb — from initial exploration to booking.
  • Multi-task learning: Multiple objectives along the journey are learned jointly within a shared model, rather than training separate single-objective rankers.
  • Industrial deployment: The approach was integrated into Airbnb's large-scale search ranking stack and validated with online experiments, following the applied-data-science tradition of prior Airbnb search papers (e.g., applied embedding techniques and deep learning ranking at KDD 2018–2019).
  • Context

    Most learning-to-rank systems optimize per-request metrics such as booking probability for a single query. Airbnb's journey framing reflects a broader industry trend: optimizing cumulative, multi-session outcomes (discoverability, engagement, and eventual conversion) rather than myopic per-impression objectives. Multi-task learning enables the ranker to balance short-term signals (clicks) against longer-term goals (wishlist adds, contact, booking) within one architecture.

    Practical implications

  • Architecture: Shared-bottom or gated multi-task architectures can serve multiple journey objectives in a single forward pass, controlling serving latency.
  • Data: Sequential user behavior logs (impressions, clicks, wishlists, bookings) are the core training signal; label definitions across the journey matter greatly.
  • Evaluation: Offline ranking metrics must be complemented by online A/B tests measuring journey-level and long-term outcomes.
  • > Note: This index post summarizes the paper's public metadata and known context. For exact architecture details, training recipes, and quantitative online experiment results, refer to the original PDF via the ACM Digital Library link above.

    Related reading

  • Airbnb's earlier KDD papers on embedding-based personalization and deep learning for search ranking
  • Multi-task and multi-objective ranking literature in e-commerce and ads (e.g., MMoE-style architectures, conversion prediction)

Tags

#airbnb#multi-task-learning#search-ranking#kdd-2023#learning-to-rank#personalization#recommender-systems#industrial-ml

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178209020