面向电子商务系统的多业务确定性调度：一种图神经网络增强的深度强化学习方法

Multi-Service Deterministic Scheduling for E-Commerce Systems: A Graph Neural Network-Enhanced Deep Reinforcement Learning Approach

摘要: 在现代电子商务系统中，业务请求通常呈现出显著的异构时间特性。不同类型的业务(如在线支付、订单确认、实时推荐与风险控制)具有差异化的服务级别协议(SLA)约束，其截止时间与时延敏感性存在显著差异。这一特性使得传统基于统一策略的请求调度方法难以同时兼顾系统效率与服务质量：一方面，统一调度容易导致高实时性业务请求错过时延约束；另一方面，过于保守的资源分配策略又会造成计算与通信资源的低效利用，增加系统运营成本。针对上述问题，本文提出了一种异构图神经网络增强的深度强化学习调度框架，面向云–边–服务节点协同的电子商务业务调度场景。该框架在构建的异构知识图谱中显式建模业务类型、业务请求与计算资源之间的关联关系，使图神经网络能够有效捕获不同业务请求之间的依赖关系、资源状态动态以及业务类型约束。在此基础上，引入一种熵正则化自适应调度策略，在满足关键业务时延与SLA约束的前提下，实现系统调度性能与资源利用率的稳定优化。

Abstract: In modern e-commerce systems, business requests typically exhibit significant heterogeneous temporal characteristics. Different types of business operations (such as online payments, order confirmation, real-time recommendations, and risk control) have differentiated service level agreement (SLA) constraints, with significant differences in deadlines and latency sensitivity. This characteristic makes it difficult for traditional request scheduling methods based on a unified strategy to simultaneously balance system efficiency and service quality: on the one hand, unified scheduling easily leads to high-real-time business requests missing latency constraints; on the other hand, overly conservative resource allocation strategies result in inefficient utilization of computing and communication resources, increasing system operating costs. To address these issues, this paper proposes a heterogeneous graph neural network-enhanced deep reinforcement learning scheduling framework for cloud-edge-service node collaborative e-commerce business scheduling scenarios. This framework explicitly models the relationships between business types, business requests, and computing resources in a constructed heterogeneous knowledge graph, enabling the graph neural network to effectively capture the dependencies between different business requests, dynamic resource states, and business type constraints. Based on this, an entropy-regularized adaptive scheduling strategy is introduced to achieve stable optimization of system scheduling performance and resource utilization while satisfying critical business latency and SLA constraints.