Selected Meituan Papers from CIKM 2024: Summaries of Eight Research Works

This article selects eight papers from Meituan's technical team that were accepted at CIKM 2024, covering self‑supervised learning, explanation generation, CTR prediction, cross‑domain recommendation, vector retrieval, image generation, timeliness prediction and other topics. The papers include both independent Meituan research and collaborations with universities and research institutes, aiming to provide helpful insights for researchers in related fields.

CIKM is a top international conference in information retrieval, knowledge management and databases. In 2024, CIKM received 1,496 submissions worldwide and accepted 347 papers, an acceptance rate of about 23%.

01

Relative Contrastive Learning for Sequential Recommendation with Similarity‑based Positive Sample Selection

Paper Type: Poster (Full Research Paper track)

Paper Link: PDF

Abstract: Sequential recommendation often suffers from severe data sparsity. Contrastive learning provides self‑supervised signals to strengthen training, but existing methods rely on data augmentation that may unintentionally alter user intent. This work proposes using similar sequences (with different target items) as additional positive samples and introduces a Relative Contrastive Learning (RCL) framework with two‑level positive sample selection and a weighted relative contrast loss. Experiments on public and Meituan datasets show RCL outperforms existing methods and has been deployed in Meituan’s homepage recommendation flow.

02

Aligning Explanations for Recommendation with Rating and Feature via Maximizing Mutual Information

Paper Type: Research Track Full Paper

Paper Link: PDF

Abstract: Natural‑language explanations can improve user satisfaction, but current generation methods often ignore consistency with predicted ratings or important item features. The authors propose the MMI (Maximum Mutual Information) framework, which uses mutual information as a consistency metric and trains a MINE‑based estimator. The estimator provides a reward for reinforcement‑learning fine‑tuning of the explanation generator, encouraging explanations that align with ratings and features. Experiments on three public datasets and user studies demonstrate significant improvements in consistency and user satisfaction.

03

Enhancing CTR Prediction through Sequential Recommendation Pre‑training: Introducing the SRP4CTR Framework

Paper Type: Short Paper

Paper Link: PDF

Abstract: Understanding user interest is crucial for CTR prediction. Existing methods integrate pre‑trained recommendation models into downstream tasks but overlook inference cost and efficient knowledge transfer. This paper presents the SRP4CTR framework, which discusses the impact of pre‑training on inference, proposes a new pre‑training method preserving information completeness, introduces a cross‑attention module for low‑cost bridging, and a self‑query technique for knowledge transfer. Offline and online experiments show the proposed method outperforms prior baselines.

04

EXIT: An EXplicit Interest Transfer Framework for Cross‑Domain Recommendation

Paper Type: Applied Research Paper

Paper Link: PDF

Abstract: Cross‑domain recommendation aims to leverage knowledge from other domains to improve prediction accuracy. Existing implicit methods ignore differences in service functions and item presentation across domains, leading to negative transfer. This work proposes an explicit interest transfer framework that models the probability of source‑domain signals transferring to the target domain under different user contexts, enabling selective signal filtering without complex network structures. The algorithm has been deployed in Meituan’s homepage recommendation system.

05

VIER: Visual Imagination Enhanced Retrieval in Sponsored Search

Paper Type: Short Paper

Paper Link: PDF

Abstract: Vector retrieval is essential for search systems, but queries in instant retail can be extremely short or noisy. Users often have visual expectations for the desired product. The authors propose a multimodal retrieval model that reconstructs image representations of both common and personalized query aspects, fusing them with semantic and behavioral features. Online A/B tests in a sponsored search system show significant gains in revenue, clicks, and CTR. Reviewers note this as one of the first papers on multimodal‑enhanced search understanding.

06

Design Element Aware Poster Layout Generation

Paper Type: Full Research Paper

Paper Link: PDF

Abstract: Poster layout generation has progressed, yet most methods ignore design elements such as text, logos, and patterns, leading to visual defects. This work defines a new task—design‑element‑aware poster layout generation—and proposes the Design Element aware Transformer (DET), an encoder‑decoder network that extracts multi‑scale background features via deformable self‑attention and aligns them with element features through deformable cross‑attention. A new metric AspDiff measures layout‑element matching. Experiments on three public datasets show DET produces layouts better aligned with design elements, and the method has been applied in Meituan’s advertising scenarios.

07

Process‑Informed Deep Learning for Enhanced Order Fulfillment Cycle Time Prediction in On‑Demand Grocery Retailing

Paper Type: Applied Research Paper

Paper Link: PDF

Abstract: Accurate prediction of Order Fulfillment Cycle Time (OFCT) is critical for on‑demand grocery retail. The authors model the distinct operational characteristics of OGR compared to food delivery and propose a deep learning model that explicitly incorporates order volume, production capacity, delivery capacity, and dispatch strategies. Multiple RNN modules evaluate dynamic load, while attention mechanisms capture interactions among orders and riders. A deep Bayesian multi‑task learning component (DBMTL) captures upstream‑downstream effects. Experiments on Meituan’s OGR dataset demonstrate significant performance improvements.

08

Collaborative Scope: Encountering the Substitution Effect within the Delivery Scope in Online Food Delivery Platform

Paper Type: Applied Research Paper

Paper Link: PDF

Abstract: The delivery scope defines the geographic area a merchant serves, influencing the set of merchants visible to users. Existing conversion models ignore the substitution relationship among merchants, leading to inaccurate predictions. This paper models the problem as a multi‑merchant selection task and proposes a machine‑learning plus combinatorial‑optimization framework that estimates order changes from the user’s perspective while respecting substitution priors. A first‑order Taylor approximation improves solving efficiency. The framework has been fully deployed in Meituan’s “Pin Hao Fan” business, reducing delivery distance and improving efficiency without sacrificing scale.

Meituan’s research cooperation platform aims to bridge Meituan’s technical teams with universities, research institutes, and think tanks, leveraging rich business scenarios and data to explore frontier technologies such as robotics, AI, big data, IoT, autonomous driving, and optimization. Interested parties can contact: [email protected] .