Emotion recognition using electroencephalography (EEG) represents a significant area of study in brain-machine interfaces. To address this multifaceted challenge, it is crucial to improve the ability of EEG features to represent emotional states. A hypergraph-based methodology allows for the depiction of higher-order spatial correlations to develop distinguishing emotional features. However, the original hypergraph may lack robustness due to potential interference among local channels. In addition, excessively coarse hypergraph granularity can result in the loss of critical information. To mitigate these issues, we propose hypergraph group learning, which aims to balance robustness with the retention of detailed information. In this study, we model temporal and spatial dependencies across varying granularities using Hypergraph Group Learning to achieve a discriminative representation of emotional features. We used multiple CNN convolutions to map EEG signals from different brain regions and time segments into a unified distribution. The multi-granularity hypergraph convolutional network (MGHGCN) is specifically designed to capture long-term temporal correlations among channels effectively. By integrating multiview fusion, we significantly improved the accuracy and robustness of EEG-based emotion recognition. Experimental results from publicly available datasets, including SEED, SEED-IV, and EMOT, validate the effectiveness of our approach, achieving precisions of 98.51 (2.46) %, 89.20 (6.13) % and 97.79 (1.31) %, respectively. These results demonstrate that our hypergraph effectively maintains both robustness and detailed information.