<p>Graph learning has become one of the most critical techniques in the area of point-of-interest (POI) recommendation, due to its excellent performance in informative representations via aggregating the embeddings of neighbors. Nevertheless, prior works either overlook the entanglement of the diverse influences stemming from different aspects of check-ins or could not explicitly present aspect-specific visiting preferences in disentangled representations, causing that the diversified POI recommendation remains a challenge. In the paper, we propose a time-aware disentangled graph contrastive learning method (T-DisenGCL), which leverages disentangled representations to directly model diverse aspects via the interactions under different time intervals to represent the user check-in preference more precisely and further achieve diversified POI recommendations. Specifically, we first model diverse aspect-specific check-in preferences at different time slots by graph learning to completely exhibit distinguishing visiting behaviors of the user in a day, and then we employ contrastive learning with representation augmentation to further enrich the informativeness of each aspect. Second, we aggregate the diverse aspects for the disentangled representation of the user, while maintaining the different components in the representation independent. Finally, the diversified POIs are recommended with element-wise vector multiplication and loss optimization. Extensive experiments on two real-world datasets demonstrate that our proposed T-DisenGCL outperforms other state-of-the-art models, and it achieves a greater balance between accuracy and diversity. Besides, T-DisenGCL lights on the potential correlations between visiting time and diversified preferences, which could benefit both users and businesses.</p>