Highlights of Meituan’s KDD 2022 Papers

Paper 01 – Mask and Reason: Pre‑Training Knowledge Graph Transformers for Complex Logical Queries

Problem: Existing KG‑embedding reasoners cannot efficiently handle complex logical queries on knowledge graphs.

Method: Introduces KG‑Transformer, a graph neural network reasoner built on transformer architecture. The model is pre‑trained on large‑scale KG data and fine‑tuned for logical query answering.

Results: Achieves state‑of‑the‑art performance on two major KG reasoning benchmarks and demonstrates strong generalization on out‑of‑domain tasks.

Download URL: https://keg.cs.tsinghua.edu.cn/jietang/publications/KDD22-Liu-et-al-KG-Transformer.pdf

Paper 02 – AutoFAS: Automatic Feature and Architecture Selection for Pre‑Ranking System

Problem: In industrial search pipelines, pre‑ranking must satisfy massive scoring volume under strict latency constraints. Prior approaches either ignore latency as an optimization variable or manually distill ranking‑stage knowledge into a fixed pre‑ranking architecture, leading to sub‑optimal performance.

Method: Formulates a Neural Architecture Search (NAS) problem that jointly optimizes (1) latency budget, (2) guidance from ranking‑stage knowledge, and (3) feature‑architecture selection. The AutoFAS framework searches over feature subsets and dual‑tower architectures under the given constraints.

Results: Delivers SOTA accuracy in Meituan’s main search scenario while respecting latency limits.

Download URL: https://dl.acm.org/doi/pdf/10.1145/3534678.3539083

Paper 03 – AutoIntent: Automatically Discovering User Consumption Intents

Problem: User consumption intents are latent drivers of behavior but are only partially observable in real‑world data, making intent discovery challenging.

Method: Proposes a hypergraph neural network with three paired hypergraphs that capture (1) time‑related, (2) location‑related, and (3) intrinsic preference relations. The disentangled intent encoder learns intent embeddings; the intent discovery decoder generates pseudo‑labels for unknown intents via semi‑supervised learning.

Results: Experiments on a large‑scale industrial dataset show >15% performance improvement over the strongest baselines.

Download URL: https://dl.acm.org/doi/pdf/10.1145/3534678.3539122

Paper 04 – Modeling Persuasion Factor of User Decision for Recommendation

Problem: Conventional recommendation models treat persuasive factors (e.g., brand, price) as black‑box inputs, limiting interpretability and performance.

Method: Constructs a heterogeneous user‑item interaction graph where persuasive copy texts form distinct edge types. A multi‑layer graph convolutional network learns representations for users, items, and copy texts. The model adaptively captures each user’s sensitivity to persuasive copy and employs counterfactual data augmentation to alleviate sparsity.

Results: Empirical evaluation on Meituan’s dataset yields significant accuracy gains and provides interpretable insights into factor influence.

Download URL: https://dl.acm.org/doi/pdf/10.1145/3534678.3539114

Paper 05 – Practical Counterfactual Policy Learning for Top‑K Recommendations

Problem: Training recommendation models from logged feedback suffers from selection bias. Existing counterfactual methods for Top‑K ranking face importance‑weight explosion, high variance, and low training efficiency.

Method: Focuses on Policy Learning for large‑scale Top‑K ranking. Proposes a practical learning framework that mitigates importance‑weight explosion and variance, and improves training efficiency. An open‑source experimental suite validates the framework.

Results: Demonstrates effective and efficient policy learning on benchmark datasets.

Download URL: https://www.csie.ntu.edu.tw/%7Ecjlin/papers/counterfactual_topk/xcf.pdf

Paper 06 – Applying Deep Learning Based Probabilistic Forecasting to Food Preparation Time for On‑Demand Delivery

Problem: Accurate estimation of restaurant food preparation time is hindered by partially observed labels (coarse time windows) and high uncertainty.

Method: Introduces a non‑parametric deep‑learning model that treats preparation time as a probabilistic variable. Designs an S‑QL loss function mathematically linked to S‑CRPS and discretizes S‑CRPS for quantile optimization.

Results: Offline evaluations and online A/B tests on Meituan’s delivery platform confirm superior performance; the model is deployed in multiple core modules of the real‑time delivery system.

Download URL: https://dl.acm.org/doi/pdf/10.1145/3534678.3539035

Paper 07 – A Framework for Multi‑Stage Bonus Allocation in Meal Delivery Platform

Problem: Unaccepted orders lead to cancellations, harming user experience. Incentive bonuses must be allocated sequentially, as each stage’s incentive influences future order survival probabilities.

Method: (1) Builds a semi‑parametric model of order completion and cancellation probabilities. (2) Applies a Lagrangian‑dual dynamic programming algorithm to compute stage‑wise multipliers from historical data. (3) Deploys an online real‑time allocation algorithm that uses the offline multipliers to determine incentive amounts per order.

Results: A/B experiments show a 25% reduction in cancelled orders compared with the baseline.

Download URL: https://dl.acm.org/doi/pdf/10.1145/3534678.3539202