Technical Solutions for Meal Combo Recommendation in Food Delivery

This article presents the technical solutions for meal‑combo recommendation used by Meituan Waimai. It first outlines the background: users face long decision times when ordering food, and merchants have weak ability and willingness to create combos, leading to low coverage of combo‑related applications.

The business goal is to generate high‑quality candidate combos for food merchants to feed downstream services such as "Today’s Combo Recommendation" and "Discount‑Tool Combo". Challenges include diverse business scenarios, non‑standard food items varying across merchants, and difficulty in evaluating combo quality.

To address these, an offline‑real‑time hybrid framework is proposed. Offline methods pre‑compute combo candidates using rule‑based and model‑based approaches, while real‑time methods generate combos on‑the‑fly for scenarios with strict constraints.

Combo Modeling: Three models are described:

Graph‑label induction model: builds a food knowledge graph from menus, recipes, and descriptions, then derives combo templates via high‑frequency order aggregation and template generalization.

Encoder‑Decoder model: treats combo generation as a sequence‑to‑sequence task. An LSTM‑based encoder extracts semantic features of dishes (name, tags, price, sales), and a Pointer Network decoder selects items from the merchant’s candidate list, using beam search for diversity.

Attention‑based model: replaces the sequential LSTM encoder with hierarchical multi‑head attention to capture unordered menu items and long‑range dependencies, and incorporates user preferences and price constraints in the decoder.

Model optimization includes teacher‑forcing training on historical orders, reinforcement learning with Monte‑Carlo sampled combos and quality rewards, and constraint‑aware pruning (e.g., price limits).

Quality Assessment: A combo quality classification model predicts four ordered grades (poor → good) using dish representations, global‑attention, and auxiliary features. The model is trained with a pair‑hinge loss and enhanced by a pre‑training task that masks a dish in a real combo and asks a Transformer to reconstruct it.

The system has been deployed in multiple Meituan Waimai products, significantly improving combo coverage and user experience. Future work aims to enrich the food knowledge graph with multimodal data and to explore scenario‑aware combos (e.g., weather‑driven, festival‑specific).