<p>With the growth of social networks and complex network data, Advanced Persistent Threat (APT) attacks have posed significant challenges to cybersecurity due to their stealthy and sophisticated nature. APT attacks typically involve a large number of interrelated entities and relationships, forming complex network structures. However, traditional methods often exhibit limited generalization capabilities when handling such multi-relational data. Additionally, the collection and annotation of APT data are labor-intensive and difficult, which hinders the model’s ability to fully leverage the data during training, ultimately affecting performance. To address these two issues, this paper proposes a semi-supervised classification method based on an Adaptive Aggregation Relational Graph Convolutional Network (RGCN) for APT attack detection. By leveraging the RGCN’s ability to capture complex relational information and employing an adaptive aggregation strategy, the proposed method effectively handles large-scale graph data while preserving critical structural features and optimizing computational efficiency. Meanwhile, through a semi-supervised learning approach, the model utilizes a collaborative learning mechanism between a small amount of labeled data and a large amount of unlabeled data. This reduces the reliance on extensive expert knowledge and manual intervention, enabling efficient knowledge discovery and model optimization with limited expert involvement. Experimental results demonstrate that the proposed model offers certain advantages in classification accuracy and generalization capability compared to other methods.</p>