In order to solve the issue of poor generalization ability of traditional classification models, an innovative meta-learning method was proposed in this paper. Firstly, a dynamic task-driven module was employed, which enabled the classification model to mine external task information; Secondly, a collaborative training method was adopted to overcome the increased complexity of calculation caused by the introduction of task-driven module; Finally, in order to exploit the relevant information between the samples, the graph neural network was used to be the classification module to achieve the purpose of further improving the classification accuracy. Compared with baselines, the accuracy of this method has increased by 6.81% and 6.20% on the MiniImageNet and CIFAR-FS datasets respectively. The experimental results show that the method in this paper can effectively improve the generalization ability of the classification model.