Tourism Recommendation System: Strategy Iterations, Architecture, and Future Challenges

Background – Vacation services are a key part of online travel. User interests are vague and vary with season, weather, and user attributes, making traditional information retrieval insufficient. A dedicated tourism recommendation system is needed.

Challenges – (1) Large cross‑city demand (over 30% of orders are from users visiting a city different from their residence). (2) Diverse recommendation formats (scenic spots, group tours, package deals). (3) Strong seasonality. (4) Highly personalized needs (e.g., families vs. couples).

Recall Strategy Iteration – Started in 2015 with a simple hot‑sale strategy based on the user's resident city. Introduced decay‑based scoring, POI‑level recall, handling of cross‑city (local vs. non‑local) orders, time‑context filtering, and strong‑correlation strategies using recent user behavior. Added location‑based recall using real‑time GPS, collaborative filtering (ItemCF and UserCF), and similarity improvements.

Ranking Strategy Iteration – Early stage used hot‑sale scores directly. Later introduced a Rerank layer using XGBoost with features such as sales, ratings, price, refunds, contextual features (hour of day, day of week, city ID), distance, and user‑POI interaction signals. Sample balancing (1:10 positive‑negative) and model training on Spark/Scala were employed. Subsequent enhancements added short‑term and long‑term models, weather/context features, and richer user‑POI features.

System Architecture – Consists of four layers: offline computation (ETL, Spark, Storm), core data layer (Elasticsearch, DataHub, Tair), recommendation service layer (recall, filtering, Rerank, post‑Rerank), and application layer (multiple front‑end scenarios). Monitoring uses Falcon for JVM/ES/业务 metrics and Hystrix for circuit‑breakers.

Application Scenarios – Supports ~20 scenarios including group‑tour recommendation, cross‑city filtering, tag‑based discovery, search‑with‑few‑or‑no‑results recommendation, and hotel‑tourism cross‑recommendation.

Future Challenges – Expand recall breadth (matrix factorization, graph mining) and depth (LLR, multi‑time/user similarity). Enhance ranking with DNNs, richer user/context features, and better evaluation metrics. Refactor engineering for lighter architecture and add new scenarios such as post‑purchase recommendation and multimodal travel (flights, trains).

Author – Zheng Gang, Senior Technical Expert at Meituan‑Dianping, focusing on search, ranking, and recommendation for hotel‑travel.