Novelty Recommendation for Meituan Food Delivery: System Design, Challenges, and Solutions

Background

In the food‑delivery scenario, users have strong repeat‑purchase behavior and high order frequency. While they often want to order from familiar merchants, they also seek new tastes. Since 2022, Meituan’s recommendation team has built a systematic solution for novelty recommendation on the home‑feed.

Key metrics show cumulative improvements: exposure novelty +19%, novelty positive‑rating +7%, and novelty bad‑case rate –18%.

1. Why Optimize User Novelty?

Historical optimization focused on transaction efficiency (RPM) and GMV, neglecting users’ desire for new experiences. This led to repeated exposure of the same merchants, creating an "information bubble" and degrading user experience.

2. Importance of Novelty in Food‑Delivery

Unlike e‑commerce where candidate pools are huge and simple filtering of already‑seen items suffices, food‑delivery candidates are limited (hundreds to thousands) and users order frequently. Therefore, a strategy that merely filters known merchants is neither realistic nor effective.

3. Definition of User Novelty

Novelty is defined as merchants that a user has not seen or ordered within a recent window (e.g., not seen in the last 7 days, not clicked in 30 days, or not ordered in 90 days) but is willing to purchase if recommended.

4. Evaluation System

The core metric is exposure novelty@Top N (the proportion of novel merchants in the first N items, N=10). Supporting metrics include click novelty, order novelty, novelty positive‑rating (≥4 stars), and novelty bad‑case rate (user‑reported negative experience).

5. Solution Overview

The feed pipeline consists of Recall → Precise Ranking → Mixed Ranking. Optimizations were applied at each stage:

Recall: Replaced Word2Vec similarity with dual‑tower merchant embeddings, added an I2I novelty recall path, and introduced a GCN‑based side‑information recall to capture high‑order user‑merchant relations.

Precise Ranking: Added novelty‑related features (negative feedback, long‑sequence signals, distance, mealtime, tag ID) and a novelty‑aware loss term to improve estimation of novel merchants while keeping UV_RPM stable.

Mixed Ranking: Implemented a weighted formula (a·pCTR + b·pCXR + c·novelty share) and an Evolution‑Strategy (ES) model for personalized hyper‑parameter tuning, enabling “thousand‑person‑thousand‑face” sorting.

Dynamic Insertion: Designed a reinforcement‑learning based insertion policy (Thompson sampling, D3QN) to decide where to place novel merchants, balancing exposure novelty with RPM constraints.

Interactive Recommendation: Added real‑time interaction signals to dynamically insert novel merchants after a user visits a merchant detail page.

6. Experimental Results

Online A/B tests (UV_RPM drop ≤0.5%) yielded:

Exposure novelty +19% overall, +8% after full system rollout.

Order novelty +25%.

Novelty positive‑rating +7% and bad‑case rate –18%.

Average user depth +1% and exposure count +1.5%.

7. Future Work

Iterate the definition of novel merchants to better align with purchase intent.

Apply causal inference and multimodal features to further improve ranking generalization.

Explore richer interactive formats (video, re‑ranking) for enhanced user experience.