Hotel Search Relevance Construction and Modeling at Fliggy (Alibaba)

Introduction The talk, delivered by Alibaba algorithm expert Lin Rui and edited by Li Peng from Chongqing University of Posts and Telecommunications, introduces the construction of hotel search relevance for Fliggy, focusing on providing users with fast, personalized hotel results.

Hotel Search Background Hotel search is a major vertical entry on the Fliggy app, also accessible via Taobao and Alipay. Unlike traditional text‑only search, hotel queries involve multiple intents, rich filters (price, star level, location), and sparse user behavior, making personalization and relevance challenging.

Search Architecture The system follows a classic pipeline: a Search Processor (SP) receives user requests, parses queries (QP), performs an initial ranking, and then a TPP service refines the ranking using offline user behavior, real‑time interaction features, and hotel attributes such as name, location, recent transactions, and inventory.

Relevance Overview

1. Scenarios and Relevance Different scenarios (nearby search, POI/landmark search, brand name search) emphasize distance, textual match, or a combination of both. Users may also have periodic needs (e.g., business trips), requiring strong personalization.

2. Textual Relevance A two‑stage approach is used: coarse ranking with BM25/Jaccard to filter candidates, followed by a fine‑ranking model that extracts transformer‑based query and title embeddings, computes element‑wise differences and products, and feeds them into a feed‑forward network as similarity features.

3. Spatial Relevance Initial features include user‑to‑hotel distance and POI‑to‑hotel distance. To improve spatial modeling, geohash codes are converted to binary strings, treated as textual tokens, and encoded into vectors that capture geographic proximity more effectively than raw distance alone.

4. Multi‑Scenario Relevance A multi‑task MLP architecture processes both generic features and scenario‑specific features (e.g., nearby, business district, brand search) to produce unified representations that handle diverse user intents.

5. Detail‑Page Signals Since clicks and transactions are sparse, user actions on the hotel detail page (viewing price, room type, membership info) are used as auxiliary labels to enrich training data and improve relevance learning.

Future Work Planned improvements include better spatial‑price estimation, two‑dimensional distance modeling, city‑level price proportion estimation, upgrading spatial‑textual relevance models with more complex online models, and incorporating historical search sequences for contextual relevance.

Conclusion The presentation wraps up with thanks and a call for audience engagement.