Intelligent Companion Search Guidance for Meituan Waimai: Challenges, Solutions, and Future Directions

Compared with other e‑commerce scenarios, the food‑delivery (waimai) scenario demands a higher real‑time capability to discover and respond to user interests. Meituan’s delivery algorithm team has built an intelligent companion‑type guidance architecture for the waimai search scene, achieving significant improvements.

Background : Search guidance (also called search recommendation) provides personalized query suggestions at various points in the search flow (in‑box words, below‑box words, discovery module, etc.). The goal is to help users find merchants and dishes more efficiently.

Challenges include transforming passive guidance into proactive, real‑time assistance; aligning guidance optimization with overall conversion goals; and handling data sparsity for new or small scenarios.

Main Work is divided into two parts:

1. Smart Refresh of In‑Box Words – a 4W1H framework (where/what/how, who/when) was used to design a query‑recommendation pipeline that triggers refresh based on Alita edge‑intelligence intent perception, real‑time user signals, and session‑level features. The pipeline collects user behavior on the client, passes intent signals to the backend, and returns refreshed queries. Real‑time trigger conditions (e.g., store click or stay >2 s) were selected to balance exposure and relevance.

2. Unified Multi‑Scenario Query Recommendation Model – a unified model covering five guidance modules (in‑box, below‑box, discovery, smart refresh, “you may also search”) was built. It includes feature construction (User, POI, Query, Context, and combined dimensions), sample construction with invalid‑sample filtering, multi‑objective optimization (CTR + CXR), and a self‑supervised pre‑training + fine‑tuning scheme. Pre‑training predicts the next user‑entered query (Next‑Query Prediction) using historical behavior and SUG data; fine‑tuning incorporates implicit feedback from all scenarios.

Key Techniques :

Real‑time edge‑intelligence (Alita) for intent detection.

Heterogeneous behavior sequence modeling with DIN attention and side‑info features.

Self‑supervised pre‑training (NSP/NWP‑style) and multi‑task fine‑tuning.

Multi‑objective Share‑Bottom architecture combining CTR and CXR.

Results : Online experiments show +1.02 % DAU, +1.70 % UV_RPM, and large lifts in UV_CTR for in‑box (+187 %), below‑box (+141 %), and discovery (+121 %). The unified model improves both in‑scenario CTR and global click/conversion metrics without degrading ranking quality.

Future Directions include extending smart refresh to more pages (feed, result page), leveraging large‑model pre‑training for query embeddings, and further breaking the recommendation feedback loop.