Reproduce “GenMatch: An End-to-End Generative Matching Framework for Micro-View Order-Dispatching in Ride-Hailing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chuang Liu, Yuxueqing Zhang, Tengfei Lyu, Zirui Yuan, Weiqi Hu, Yanghan Cheng, Ming Wang, Li Ma and 1 more
Researchers at Didi Chuxing and The Hong Kong University of Science and Technology (Guangzhou) developed GenMatch, an end-to-end generative framework for micro-view order-dispatching in ride-hailing, overcoming the limitations of conventional multi-stage systems. Deployed on a major industrial platform, GenMatch achieved a +2.97% increase in Gross Merchandise Volume, a +2.26% Answer Ratio, and a +3.86% Completion Ratio in online A/B tests.
Key insights
- Formulating MICOD as an end-to-end generative task effectively addresses cross-stage objective inconsistency by optimizing directly for the final batch-level assignment.
- Explicitly modeling interdependencies and competition within a bipartite graph through a context-aware encoder is crucial for effective batch-level representation.
- Learning a unified business utility from diverse feedback signals, combined with a state-aware pointer mechanism, enables the generative model to make coordinated, feasible, and business-optimal decisions.
- Dynamic, state-aware sequential generation significantly outperforms static, one-shot optimization methods, as demonstrated by the performance gap between full GenMatch and variants using a fixed solver.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.