Selected Meituan Technical Papers from KDD 2023: Summaries of Seven Research Works

This article presents a curated collection of seven papers authored by the Meituan technology team that were accepted at KDD 2023. The papers span topics such as feed‑stream recommendation, cross‑domain recommendation, takeaway recommendation, online bonus allocation, contour‑based instance segmentation, living‑needs prediction, and multilingual recommendation.

ACM SIGKDD (the International Conference on Knowledge Discovery and Data Mining) is a premier venue for researchers and practitioners to share advances in data mining and knowledge discovery. Since its first edition in 1995, SIGKDD has become a key platform for presenting cutting‑edge research and fostering collaboration.

01. "PIER: Permutation‑Level Interest‑Based End‑to‑End Re‑ranking Framework in E‑commerce" Authors: Xiao Wen, Yang Fan (co‑first), Wang Ze, Xiao Xu, Guo Gang, Yong Kang, Xing Xing, Wang Dong, et al. (Meituan) Type: Full paper Download: PDF Abstract: Feed‑stream recommendation is a dominant paradigm whose quality directly impacts user experience, merchant revenue, and platform income. Traditional point‑CTR prediction ignores contextual information from surrounding items, leading to sub‑optimal performance. Existing re‑ranking approaches either consider only position bias and upstream context or use Listwise estimation followed by re‑ranking, which suffers from evaluation‑before‑re‑ranking issues. PIER proposes a two‑stage end‑to‑end re‑ranking framework comprising FPSM (Fast Permutation‑Level Candidate Generation via SimHash) and OCPM (Omni‑Contextual Preference Modeling with a novel full‑attention mechanism). A contrastive learning loss jointly trains both modules, and OCPM predictions guide FPSM to generate better lists. Offline experiments on public and industrial datasets show PIER outperforms baselines, and the model has been deployed in Meituan’s delivery advertising scenario with significant gains.

02. "A Collaborative Transfer Learning Framework for Cross‑domain Recommendation" Authors: Zhang Wei, Peng Ye, Zhang Bo, Xing Xing, Wang Dong, et al. (Meituan) Type: Full paper Download: PDF Abstract: Different business domains exhibit heterogeneous CTR distributions, making cross‑domain knowledge transfer challenging. Naïve pre‑training + fine‑tuning can cause negative transfer, while mixed‑sample multi‑task training suffers from a seesaw effect. The proposed CCTL framework evaluates the information gain of source‑domain samples for the target domain, assigns adaptive weights, and incorporates contrastive learning (REN) and token‑level alignment (SAN) to preserve domain‑specific differences. The system has been deployed in Meituan’s take‑away advertising and yields noticeable improvements.

03. "Modeling Dual Period‑Varying Preferences for Takeaway Recommendation" Authors: Yu Ting (Meituan), Wu Yiqing (Institute of Computing Technology, CAS), Zhu Yongchun (Institute of Computing Technology, CAS), Zhuang Fuzhen (Beihang University), Rui Dong (Meituan), Bei Hai (Meituan), Zhan Bo (Meituan), An Zhulin (Institute of Computing Technology, CAS), Xu Yongjun (Institute of Computing Technology, CAS) Type: Full paper Download: PDF Abstract: Takeaway recommendation must capture (1) dual interaction preferences (user‑shop and user‑dish) and (2) intra‑day preference shifts. The DPVP model introduces a Dual‑Interaction‑Aware module, a Time‑based Decomposition module, and a Time‑Aware Gate to model these aspects. Both offline and online experiments demonstrate superior performance over state‑of‑the‑art baselines, and the model is now live on Meituan’s platform.

04. "A Multi‑stage Framework for Online Bonus Allocation Based on Constrained User Intent Detection" Authors: Wang Chao, Xiao Wei, Xu Shuai, Wang Zhe, Zhi Qiang, Feng Yan, You An, Chen Yu, et al. (Meituan) Type: Full paper Download: PDF Abstract: Bonus allocation is modeled as a knapsack optimization problem with a two‑stage pipeline: (1) user‑intent detection (deep representation learning with monotonic constraints) and (2) online allocation (convex‑constrained optimization). A feedback‑control module mitigates budget overruns. Experiments on real Meituan payment data and online A/B tests confirm the framework’s effectiveness.

05. "C‑AOI: Contour‑based Instance Segmentation for High‑Quality Areas‑of‑Interest in Online Food Delivery Platform" Authors: Yi Da, Li Ying, Da Ping, Shui Ping, Fang Xiao, Jing Hua, Ren Qing, Zhi Zhao, et al. (Meituan) Type: Full paper Download: PDF Abstract: Areas‑of‑Interest (AOI) are essential for logistics decision‑making. Existing AOI generation methods rely on predefined shapes, density clustering, or post‑processing semantic segmentation, limiting contour quality. C‑AOI reformulates AOI generation as an instance‑segmentation contour regression problem, introducing a Contour Transformer, cyclic positional encoding, and adaptive matching loss. Experiments on Meituan’s delivery dataset show substantial quality gains and fast inference; the model is already in production.

06. "NEON: Living Needs Prediction System in Meituan" Authors: Lan Xiaochong (Tsinghua), Gao Chen (Tsinghua), Shi Qi (Meituan), Xiu Qi (Meituan), Ying Ge (Meituan), Zhang Han (Meituan), Hua Zhou (Meituan), Heng Liang (Meituan), Li Yong (Tsinghua) Type: Full paper Download: PDF Abstract: User intent prediction (e.g., dining, accommodation, entertainment) is crucial for Meituan’s recommendation and marketing. NEON addresses two challenges: (1) complex, multi‑factor user demand and (2) heterogeneous intent manifestation across the app. The system comprises three stages—feature mining (contextual consumption scenarios), feature fusion (user‑individual and group‑behavior networks), and multi‑task prediction (intent plus delivery‑mode preference). Large‑scale online A/B tests demonstrate significant business impact.

07. "A Hybrid Approach of Statistics and Embeddings for Multilingual and Multi‑Locale Recommendation" Authors: Wei Jia (Meituan), Zhan Jin (DataRobot), Huang Zhongshan (Freelance), Wang Lu (Microsoft), Wang Qiang (Meituan) Type: Workshop paper Download: PDF Abstract: To promote multilingual recommendation research, Amazon released a multilingual, multi‑locale shopping session dataset and organized the KDD Cup 2023 multilingual session recommendation challenge. Meituan’s solution follows a two‑stage pipeline: diverse recall strategies (I2I via co‑visit matrix, GraphEmbedding, TextTransformer, BPR) followed by a ranking model that ensembles two GBDT models using rich statistical features and embedding similarity. The approach achieved 3rd place in Track 2 and 4th in Track 1.

The article concludes by inviting academic collaborators to join Meituan’s research cooperation program, covering areas such as robotics, AI, big data, IoT, autonomous driving, and optimization. Interested parties can contact [email protected].