Optimizing the Efficiency of Heterogeneous Fusion of Multi-Source Information Nodes in Smart Government Platforms Using GraphSAGE

Abstract

To address low fusion efficiency in multi-source heterogeneous data and the difficulty of processing large-scale graph data in smart government platforms, this paper proposes a GraphSAGE-based optimization method. A heterogeneous graph structure with multiple node and edge types is first constructed, abstracting government entities into heterogeneous nodes and distinguishing business-related edge semantics. The GraphSAGE framework is improved using type-aware neighbor sampling and a multi-head attention aggregation function to preserve heterogeneous semantic features. Considering the temporal characteristics of government data, an LSTM dynamic modeling module is introduced to enhance the temporal evolution of node representations. A graph partitioning strategy combined with distributed training is further used to improve large-scale computation by optimizing communication and synchronization processes. Experiments show that the proposed method achieves node embedding accuracy of at least 0.95 and NMI of 0.88 in complex government graph scenarios, indicating high fusion quality. For 100,000 nodes, the training time is 12 seconds, with convergence achieved in 140 rounds. The results demonstrate that the method can effectively support large-scale heterogeneous graph representation, dynamic information fusion, and efficient decision-support modeling in smart platform environments.