Generative graph models have recently emerged as a powerful paradigm in self-supervised learning (SSL) for graph representation tasks, demonstrating remarkable potential in capturing complex structural and feature information. Despite their promise, most existing generative graph representation learning (GRL) methods treat the entire graph as a monolithic entity, employing random masking strategies across the entire structure. This approach often neglects the intricate interplay between different substructures of the graph, leading to entangled representations. Such entanglement can reduce the robustness of learned models and limit their ability to produce reliable outputs. To address these limitations, this article presents disentangled generative graph representation learning (DiGGR), a novel SSL framework designed to disentangle latent factors in graph representations. DiGGR introduces a disentanglement-driven approach to guide graph mask modeling, enabling the framework to learn distinct latent factors that better capture the underlying structure of graphs. By incorporating these latent factors into an end-to-end joint learning process, DiGGR enhances the clarity and structural alignment of the learned representations, while also improving their robustness and generalization across tasks. The effectiveness of DiGGR is validated through extensive experiments on 15 public datasets spanning three distinct graph learning tasks. The results consistently demonstrate that DiGGR outperforms a wide range of previous self-supervised methods, highlighting its superior ability to learn meaningful, disentangled representations. These findings underscore the potential of DiGGR to advance the state of the art in generative GRL research.