GenMatch Boosts Ride-Hailing Order Dispatch Efficiency

In ride-hailing platforms, efficiently assigning available drivers to passenger orders within each dispatch batch, known as micro-view order-dispatching, is crucial for both service quality and operational efficiency. Traditional industry solutions typically involve a multi-stage process of model prediction, value calculation, and dispatch matching. A significant drawback of this approach is the "cross-stage objective inconsistency," where optimizing individual stages does not necessarily lead to an improved overall dispatch result, as the final quality is determined by the batch-level assignment. To overcome this, researchers have developed GenMatch, an end-to-end Generative Matching framework. This is the first such framework to be successfully deployed in a real-world production environment, specifically within DiDi's international ride-hailing markets. Applying generative modeling to this problem presented three main challenges: efficiently encoding dynamic sparse bipartite graphs for each dispatch batch, learning a unified business utility from diverse feedback to replace hand-crafted value functions, and tracking evolving matching states during assignment generation. GenMatch addresses these challenges with three core components: a Context-Aware Bipartite Encoder for structured batch-level encoding, a Business-Aware Utility Learner for unified utility optimization, and a State-Aware Pointer Decoder for dynamic assignment generation. Extensive offline evaluations and online A/B tests across five cities confirmed GenMatch's consistent improvements over competitive baselines, demonstrating its effectiveness and practicality for industrial-scale order-dispatching.