Text-attributed graphs (TAGs), which associate rich textual descriptions with each node, are widely employed to represent complex relationships among real-world textual entities. Currently, representation learning for TAGs leverages large language models (LLMs) to transform node-matched textual descriptions into node features or labels, followed by the message passing in graph neural networks (GNNs) that further improves the expressiveness of graph representation learning. Nevertheless, a simple experiment we conducted demonstrates that not all LLMs are readily compatible with GNNs. A salient finding indicates that architectural heterogeneity among LLMs manifests as substantial performance gap across diverse TAGs representation learning. Moreover, the node semantics encoded by LLMs are often misaligned with the message passing in GNNs, causing performance collapse. Motivated by this observation, we propose a novel self-supervised graph learning framework called Stage-Aware Graph Contrastive Learning (SAGCL). In particular, we propose the node-oriented mixture of experts (NodeMoE) to assign suitable candidate experts for each node. It flexibly balances the strengths of different language experts by low-rank decomposition and reparameterization strategies. Subsequently, to align the inductive biases of graph structures with the semantic perception capabilities of LLMs, the message passing in GNNs is decoupled into the feature transformation stage and the feature propagation stage. Given the two stage views, stage-aware graph contrastive learning is proposed to match the node semantics encoded by the LLM with the locally aware topological patterns within the GNN via self-supervised contrastive learning. Experiments on eight datasets and three downstream tasks demonstrate the effectiveness of SAGCL.

Advanced Persistent Threat (APT) attacks—notable for their stealth, sophistication, and potential for catastrophic damage—represent one of the most critical challenges in modern cybersecurity. However, existing detection methods relying on conventional graph models are constrained to capturing pairwise interactions between entities. This limitation hinders the accurate representation of multi-entity dependencies and multidimensional relationships inherent in APT attacks, resulting in imprecise relationship modeling that undermines detection accuracy. Furthermore, efforts to enhance relational representation through complex graph frameworks often introduce significant computational burdens, hindering real-time detection capabilities. To overcome these limitations, we introduce LOHA(Louvain-Hypergraph Masked Autoencoder), a hypergraph modelingdriven APT detection framework that leverages the Louvain algorithm to build expressive hypergraph structures with minimal computational overhead, combined with a mask autoencoder and self-supervised learning for deep semantic feature extraction to enable multi-granularity APT detection. Experimental results demonstrate that LOHA outperforms traditional methods in detection performance across three public datasets while maintaining lower computational overhead, exhibiting enhanced robustness and real-time performance.