Micro-View Order Dispatching assigns available drivers to passenger requests within each dispatch batch and is a core operation for ride‑hailing platforms. Existing industrial pipelines follow a multi‑stage flow of prediction, value calculation, and matching, but the objectives of these stages are misaligned, so improving a single stage does not guarantee better overall dispatch quality. We formalize the task as a generative matching problem and introduce GenMatch, the first end‑to‑end generative matching framework deployed in a real production environment. GenMatch faces three major challenges: (1) each batch forms a dynamic sparse bipartite graph, requiring efficient batch‑level structured encoding; (2) the handcrafted value function is replaced by a unified business utility learner that must infer utility from heterogeneous feedback; (3) directly generating assignments must track the evolving matching state because each selected order‑driver pair alters the remaining feasible candidates. To address these, GenMatch comprises a Context‑Aware Bipartite Encoder, a Business‑Aware Utility Learner, and a State‑Aware Pointer Decoder, handling graph encoding, utility learning, and state‑aware pointer decoding respectively. Extensive offline experiments and online A/B tests in five cities of DiDi’s international markets consistently outperform competitive baselines, confirming GenMatch’s effectiveness and practicality for industrial order dispatching.
Blogger's Review: GenMatch brings generative modeling into dispatching, resolves cross‑stage objective inconsistency, and demonstrates the potential of end‑to‑end learning in large‑scale real‑time systems.