Interactive Recommendation System for Food Delivery Feed

Background

Interactive recommendation, first proposed by YouTube in 2018, addresses latency and weak user interaction in traditional recommender systems. Starting in late 2021, Meituan Waimai's recommendation team explored and fully deployed an interactive recommendation framework on the homepage feed in 2022, aiming to dynamically insert merchant cards based on users' real‑time needs.

Problems and Challenges

Designing a real‑time, end‑device‑centric recommendation system that matches users' instantaneous demands.

Choosing optimization goals and metrics that reflect the overall efficiency of the homepage feed, not just a single module.

Inserting interactive cards into a single‑column merchant list without causing disruptive user experience.

Understanding users' immediate intent and leveraging the inserted card to improve conversion while minimizing negative impact on surrounding items.

Main Work

3.1 Interactive Recommendation Framework

The design follows a “4W1H” approach: Where/How the card appears, Who/When it is triggered, and What it displays. Multiple teams (product, edge‑intelligence, client, service, and recommendation engineering) collaborated to build the end‑to‑end pipeline.

3.1.2 Product Forms

Several card formats were iterated: search‑term cards, multi‑merchant aggregation cards, and single‑merchant cards. Interaction logic compared overlaying the clicked card versus inserting a new card below it to reduce perceived interference.

3.2 Evaluation Methodology and Metrics

The goal is to improve overall feed conversion efficiency. Two key dimensions are measured: coverage (page‑level exposure ratio) and matching quality (relative next‑position conversion difference). Metrics such as "exposure page ratio", "exposure PV ratio", and "exposure UV ratio" are introduced to assess coverage, while a "relative next‑position difference" metric mitigates population, position, and resource‑type biases.

3.3 User Intent Understanding

Two trigger strategies are explored. The first‑trigger strategy fires immediately after a user enters a merchant detail page; the continuous‑trigger strategy repeatedly requests recommendations as the user stays longer or performs additional actions. These strategies balance exposure volume against confidence in the trigger.

3.4 Ranking Strategy

To handle distribution shift between feed list items and single‑card interactions, the team fine‑tuned the existing feed ranking model using interactive‑card samples. The model incorporates user, merchant, context, sequence, and trigger features, passes them through an MMoE layer and three MLP layers, and outputs predicted CTR and CXR.

3.4.2 Mechanism and Exposure

A mechanism module reorders candidates based on business goals (CTR, CXR, novelty). Exposure filtering rules (duplicate merchants, pre‑order merchants, brand duplicates, blacklists, high delivery fee/distance) reduce negative feedback. The final exposure decision compares the interactive card’s predicted CXR with that of the next natural merchant; only if it is higher is the card shown.

Results and Outlook

The interactive recommendation system increased homepage feed GMV by 0.43%, novelty by 1.16%, and achieved a 132% lift in conversion rate compared with the next natural merchant. Future work includes refining product forms, supporting additional business objectives (novelty, diversity), and moving more processing to edge devices for ultra‑personalized, privacy‑preserving recommendations.