Abstract: ‍ Driven by intelligent applications， the demand for enhanced communication performance has exceeded the capabilities of existing networks. However， the limitations of existing network architectures hinder their ability to adapt dynamically to the continuously evolving application requirements and complexities of a changing network environment. This poses unprecedented challenges for traditional rule-based algorithms. In contrast， artificial intelligence models， known for their exceptional data-fitting capabilities， offer distinct advantages in addressing these challenges. By leveraging artificial intelligence models， it is possible to effectively analyze and optimize the intricate operational frameworks of wireless communication networks， thereby enabling more resilient and adaptive 6G network systems. The large number of communication signals generated by complex wireless systems contain numerous fields， yet only a limited subset significantly impacts network performance. Consequently， artificial intelligence models must identify and analyze the interrelationships among these data fields to enhance the efficiency of wireless communication networks. Existing methods for wireless data analysis often overlook the multi-source heterogeneity and temporal dynamics of such data. As an emerging artificial intelligence technology， knowledge graphs can integrate expert knowledge with data awareness to more effectively represent the correlations among data. In this study， we developed a representation learning model tailored to the spatiotemporal heterogeneous nature of the wireless data knowledge graph， referred to as the spatiotemporal heterogeneous representation learning model. Specifically， the proposed approach transforms the entities and relationships in a wireless data knowledge graph into nodes and edges of a heterogeneous graph， learning node feature representations to systematically capture the interdependencies among wireless data fields. These relationships not only reveal the physical associations between data fields but also serve as theoretical guidance for performance optimization. Furthermore， the spatiotemporal heterogeneous model constructs a feature graph module that captures the structural and attribute information of heterogeneous nodes by generating new multi-hop connected subgraphs. To effectively capture the dynamic characteristics of wireless data， the proposed model incorporates a spatiotemporal dynamic module that jointly aggregates structural and temporal information from neighboring nodes， facilitating the learning of node representation vectors that evolve over time. Compared to existing methods， the proposed spatiotemporal heterogeneous representation learning model demonstrates significant advantages in understanding wireless data field relationships and predicting performance. Its effectiveness is further validated through a throughput prediction task， demonstrating its practical value in real-world applications. This study not only addresses existing gaps in wireless data knowledge graph representation learning but also provides both theoretical contributions and practical solutions for optimizing wireless communication systems.