Interactive Recommendation System for Meituan Food Delivery: Architecture, Challenges, and Evaluation

1. Background

Interactive recommendation is a real‑time, interactive recommendation module that understands user needs and provides recommendations through interaction. It was first proposed by YouTube in 2018 and later adopted by Meituan Waimai from late 2021 to continuously explore and fully deploy on the homepage feed.

1.1 What is Interactive Recommendation?

It is an interactive real‑time recommendation product module that primarily works by understanding user demand and delivering recommendations in an interactive manner.

1.2 Why is Interactive Recommendation Needed?

The food‑delivery feed has lower user dwell time than traditional e‑commerce, making it harder to capture and respond to instant user interests. Traditional pagination limits the system’s ability to adjust recommendations in real time.

2. Problems and Challenges

Building a recommendation system that matches real‑time user needs in the food‑delivery scenario.

Choosing optimization goals and measuring effects for the whole feed.

Avoiding interference when inserting interactive cards into a single‑column list.

Understanding user intent and optimizing the matching of inserted cards.

3. Main Work

3.1 Interactive Recommendation Framework

3.1.1 Overall Flow

The system follows a "4W1H" design: Where/How (card placement and form), Who/When (triggering users and timing), What (card content and strategy).

The end‑to‑end flow starts when a user clicks a merchant card, triggers the intent‑understanding engine on the client, processes features, computes recommendations, and returns results through application service, mixing, and ranking modules.

3.1.2 Product Forms

Multiple card forms were tested: search‑term cards, multi‑merchant aggregation cards, and single‑merchant cards. Interaction logic (overlay vs. insertion) was also compared to minimize user interference.

3.2 Evaluation Methods and Metrics

The goal is to improve overall feed conversion efficiency. Two main dimensions are coverage (exposure page share) and matching (relative next‑position conversion difference). Metrics such as "exposure page share", "same‑position order increment", and "relative next‑position difference" are introduced.

3.3 User Intent Understanding

Two stages: (1) defining trigger conditions (e.g., add‑to‑cart, click, dwell time) and (2) modeling instant user intent. Real‑time, fine‑grained features from the client are leveraged for more accurate intent inference.

3.3.1 First‑Trigger Strategy

After experiments, the strategy of triggering immediately after the user enters a merchant detail page was adopted.

3.3.2 Continuous‑Trigger Strategy

During the stay on the merchant page, multiple requests are sent to update recommendations as user behavior evolves.

3.4 Recommendation Ranking Strategy

3.4.1 Recall

Two‑step recall: (1) retrieve hundreds of candidate POIs using various recall algorithms; (2) filter candidates with a similar‑category filter. An Item2Item Multi‑Trigger bypass recall is added to capture instant intent.

3.4.2 Ranking

The ranking model predicts pCTR and pCXR. Due to sample distribution differences, a fine‑tuning approach is used: the feed ranking model is fine‑tuned with interactive‑recommendation samples.

3.4.3 Mechanism

Business rules (e.g., duplicate filtering, blacklist, delivery fee constraints) are applied to adjust the order of candidates based on multiple objectives such as CTR, CXR, and novelty.

3.4.4 Exposure

The top‑1 candidate from the mechanism stage is evaluated for exposure. The card is shown only if its predicted pCXR exceeds that of surrounding cards (areas B, C, D) by a calibrated factor.

4. Summary and Outlook

The interactive recommendation system, built on client‑side intelligence, improves feed conversion (+0.43% transaction value, +1.16% novelty) and achieves a 132% lift in conversion compared to the next natural merchant.

Future work includes optimizing product forms, supporting more business goals (novelty, diversity), and moving more processing to the client to achieve "thousand‑person‑thousand‑model" personalization while protecting privacy.

5. Authors

Ji Chen, Ya Cheng, Wang Wei, Cheng Long, Jiang Fei, Wang Cong, Bei Hai, etc., from Meituan Delivery Business Group and R&D Platform.

6. References

[1] Christakopoulou K, et al. Q&R: A two‑stage approach toward interactive recommendation. KDD 2018.

[2] He X, et al. Practical lessons from predicting clicks on ads at Facebook. WSDM 2014.

[3] Gong Y, et al. EdgeRec: recommender system on edge in Mobile Taobao. CIKM 2020.

[4] Edge intelligence in Dianping search re‑ranking.

[5] Huang P‑S, et al. Learning Deep Structured Semantic Models for Web Search Using Clickthrough Data. CIKM 2013.

[6] Shi Y, et al. Collaborative Filtering beyond the User‑Item Matrix. ACM CSUR 2014.

[7] Ma J, et al. Modeling task relationships in multi‑task learning with multi‑gate mixture‑of‑experts. KDD 2018.

[8] Meituan Waimai contextual intelligent traffic distribution practice.

[9] Zhou G, et al. Deep Interest Network for CTR prediction. KDD 2018.