Evolution of Meituan Travel Search Recall Strategies

Background: Meituan‑Dianping's travel search faces challenges such as cross‑region queries, diverse user intents, and noisy data.

Evaluation metrics: The team uses purchase‑rate (visits to purchase UV), click‑through rate, no‑result rate, and user satisfaction scores to assess search quality.

Strategy iteration method: A case‑driven four‑step loop (quality evaluation, problem analysis, project development, experiment iteration) guides improvements.

Recall strategy evolution: From nationwide recall to modular result presentation, intent‑based recall, multi‑stage recall (first, second, third), and parallel task execution.

Intent classification: Queries are categorized into POI, administrative region, category, itinerary, tourism keywords, travel agency, ticket, and non‑travel intents, using chunk analysis and CRF models.

Chunk analysis: The query is segmented into chunks with BMES tags; a CRF model trained on labeled logs identifies chunk boundaries and tags.

Coarse ranking improvements: Distance segmentation, confidence‑weighted rating, new‑item sales smoothing, and multiplicative factor combination are introduced.

Text relevance enhancements: Revised BM25‑like formula reduces field length bias and uses max over fields with dynamic weights.

Term importance: A four‑level importance scheme (super important, required, important, unimportant) is learned with XGBoost using textual, statistical, language‑model, and chunk features.

Full‑field recall and final workflow: All POI fields are used for matching; the end‑to‑end recall pipeline includes preprocessing, tokenization, stop‑word removal, synonym rewrite, chunk‑based intent detection, term importance scoring, and multi‑stage recall.

Conclusion: Incremental, data‑driven iterations balance “no‑recall” and “mis‑recall”, steadily improving travel search quality.