Journal of Networking Technology Maleerat Maliyaem 17 3 2026 https://doi.org/10.6025/jnt/2026/17/3/101-135 https://www.dline.info/jnt/fulltext/v17n3/jntv17n3_1.pdf The rapid expansion of graph-structured data necessitates advanced analytical methods for understanding complex organizational communication networks. This study presents a comprehensive graph representation learning framework tailored for large scale organizational email networks, integrating deep graph neural networks (GNNs), graph autoencoders, and explainable artificial intelligence (XAI). Utilizing the ia-email- EU-dir dataset, comprising over 265,000 nodes and 420,000 directed edges, we systematically evaluate classical embedding techniques against advanced GNN architectures. Our proposed hierarchical encoderdecoder architecture, specifically the Adversarially Regularized Graph Autoencoder (ARGA), demonstrates superior performance in capturing both local connectivity and global topological dependencies. ARGA achieved the highest predictive accuracy for link prediction tasks, yielding an Area Under the Curve (AUC) of 0.948 and an Average Precision (AP) of 0.935, alongside the lowest reconstruction loss. Furthermore, the framework incorporates multiple XAI modules including GNNExplainer, GraphSHAP, and attention weight visualization to decode latent representations, thereby identifying critical communication hubs, influential subgraphs, and hidden community structures. By bridging scalable representation learning with model interpretability, this unified pipeline offers a robust solution for organizational analytics, enabling practical applications such as anomaly detection, knowledge flow mapping, and hierarchy discovery. Ultimately, the empirical findings indicate the profound efficacy of adversarially regularized probabilistic models in deciphering the intricate, evolving dynamics of sparse, scale free enterprise communication networks, paving the way for trustworthy AI in enterprise decision support and organizational behavior analysis.