Path optimization for balanced allocation of educational resources based on graph neural networks

Abstract

To address the technical bottlenecks of spatiotemporal mismatch between supply and demand in regional educational resource allocation, and the inherent tendency of traditional operational research algorithms to fall into local optima within high-dimensional networks, this paper proposes an End-to-End Balanced Scheduling Model Based on Heterogeneous Graph Attention Encoding and Proximal Policy Decoding (HGA-PPD). First, the urban educational system is modeled as a dynamic heterogeneous graph comprising multi-source entities. A multi-head graph attention encoder integrating a time decay factor is designed to accurately extract the asymmetric spatial radiation features and dynamic resource gaps between core schools and underprivileged schools. Second, to overcome the dimensional explosion of the joint action space caused by discrete routing and continuous allocation, a decoupled Proximal Policy Decoder is constructed during the reinforcement learning phase. This decoder achieves synchronous optimization of scheduling routing and allocation proportions through a dual-branch Actor network. Furthermore, a multi-objective reward mechanism balancing the Global Demand Fulfillment Ratio (GDFR) and the System Global Routing Impedance (SGRI) is introduced to guide the model towards the Pareto optimal front of fairness and efficiency. Comparative experiments on the EduGraph-CityX dataset demonstrate that, compared with classical heuristic algorithms and standard graph reinforcement baselines, the GDFR of the HGA-PPD model is improved by up to 19.3% (reaching 91.8%), while significantly reducing scheduling impedance under millisecond-level inference latency. This study provides a novel datadriven paradigm for the equitable allocation and dynamic flow of resources within the framework of smart education.