Evolution of Meituan-Dianping Search Core Ranking: From Traditional Models to LambdaDNN Listwise Deep Learning

This article introduces the evolution of the core ranking layer of Meituan-Dianping (DP) search, covering the transition from traditional machine‑learning models to large‑scale deep‑learning ranking models that integrate knowledge‑graph information. The new Listwise deep learning model, LambdaDNN, and related feature‑engineering practices and toolchains are described.

1. Introduction

DP search is the main entry for users to find information on the DP app. The diversity of user intents, business scenarios, user types, and LBS (location‑based service) requirements creates unique challenges that differ from generic web search. To address these challenges, the DP search team upgraded NLP techniques, incorporated knowledge‑graph signals, and adopted deep learning across the ranking pipeline.

2. Knowledge‑Graph‑Driven Architecture

Meituan’s "Meituan Brain" builds the world’s largest restaurant‑entertainment knowledge graph. By extracting entities and relations from user reviews, the graph provides fine‑grained tags (e.g., "good parking", "delicious food"). The knowledge graph is injected into a five‑layer search architecture, with the core ranking layer being the focus of this article.

3. Ranking Model Exploration and Practice

3.1 Traditional Machine‑Learning Models

Early DP ranking used linear models (LR), factorization machines (FM), and gradient‑boosted decision trees (GBDT). LR offers strong interpretability but requires extensive manual feature engineering. FM adds second‑order interactions automatically. GBDT captures high‑order statistical features but struggles with sparse high‑dimensional inputs.

3.2 Deep Neural Network Models

From late 2018, DP migrated to deep learning models to handle massive user behavior data and knowledge‑graph features. Advantages include strong fitting ability, superior feature representation (via embeddings), and automatic feature combination (e.g., DeepFM, DeepCrossNetwork). The team built a Wide&Deep architecture where the Wide part consumes fine‑grained statistical features and the Deep part learns embeddings for high‑dimensional categorical features.

3.3 Feature‑Engineering Practices

Key practices include feature normalization, discretization (equal‑frequency bucketing and tree‑based bucketing), feature crossing (e.g., gender × category), and extensive use of embeddings for user behavior sequences, user IDs, merchant information, and image features. Sequence embeddings are incorporated via pooling, RNN (LSTM/GRU), or attention mechanisms.

4. Listwise Deep Learning Ranking Algorithm: LambdaDNN

4.1 Gap Between Business Metrics and Model Objectives

Standard pointwise log‑loss optimizes per‑item click probability, which does not align well with business metrics such as QV_CTR or Session Success Rate (SSR) that emphasize top‑ranked results. Listwise objectives like NDCG better reflect ranking quality.

4.2 From Log‑Loss to NDCG via LambdaRank

LambdaRank constructs a special gradient (Lambda gradient) that approximates the change in NDCG when swapping two documents. By back‑propagating this gradient through a deep network, the model directly optimizes NDCG. The resulting architecture is called LambdaDNN.

4.3 Engineering Implementation

Training uses TensorFlow distributed framework. Samples are grouped by QueryId into TFRecord files to ensure that all documents of the same query are processed together. Techniques such as dynamic padding removal, multi‑hot categorical concatenation, sparse‑tensor updates, and parameter sharding across PS servers reduce memory and communication overhead.

4.4 Further Optimization of NDCG

Real‑world position bias differs from the theoretical exponential decay used in NDCG. The team fitted empirical exposure‑CTR curves and introduced a position‑bias term to better reflect mobile scrolling behavior.

4.5 Lambda‑Based Deep Ranking Frameworks

Beyond LambdaDNN, the team explored LambdaDeepFM and LambdaDCN (Deep & Cross Network). DCN adds explicit cross layers that learn feature interactions efficiently. Experiments show that Lambda‑DCN further improves NDCG.

5. Deep Ranking Diagnosis System

Because deep models are black boxes, the team built a diagnostic tool named Athena. It supports Pairwise and Listwise modes, leveraging LIME‑style perturbations to identify feature contributions for individual ranking decisions, helping engineers quickly respond to bad cases and improve model interpretability.

6. Conclusion and Outlook

The migration from tree‑based models to large‑scale deep ranking has yielded significant KPI improvements. Future directions include deeper exploitation of knowledge‑graph structure via graph embeddings, further network architecture innovations beyond fully‑connected DNNs, combined loss functions (Log‑Loss + Lambda‑Loss) with multi‑task learning, and exploration of groupwise ranking models such as TF‑Ranking.