In recent years, object detection techniques have served as a vital component in many applications, but most models rely heavily on large-scale labelled data, resulting in huge data collection and labelling costs. Therefore, research on few-shot object detection holds greater practical value. Currently, the meta-learning-based paradigm is the mainstream method. However, due to the scarcity and randomness of support class features, the model may not fully learn class features, resulting in missed or inaccurate detections. To handle this concern, we introduce a hybrid data variational graph feature aggregation model, which utilizes variational autoencoders to simulate the generation of class features that are subject to the intra-class distribution. At the same time, we utilize graph convolutional networks to deeply explore latent correlations between support and query features, aiming to obtain comprehensive and discriminative class features. In addition, this paper designs a Feature Selection Module to optimize feature interaction, improving the detector’s recognition and generalization capabilities. Experimental comparisons are conducted on PASCAL VOC and MS-COCO, which reveal that our approach exhibits superior overall average performance compared to other methods.