Abstract: In multi-label classification, an instance may have multiple labels, and in few-shot scenario, the performance of model is more vulnerable to the complex semantic features in the instance. However, current prototype network only takes the mean value of instances in support set as label prototype. Therefore, there is noise caused by features of other labels in the calculated prototype, weakening the differences among prototypes, and affecting the prediction effect. To solve the above problem, a classification method was proposed for prototype network with instance-level attention in multi-label few-shot learning. This method was designed to reduce the interference resulted from features of other labels by increasing the weight of instances with high correlation between the support set and label, to improve the discrimination among prototypes, and further to improve the accuracy of prediction. The experimental results show that the proposed method can strengthen the learning ability of multi-label prototype network, and the classification effect is significantly improved.