Abstract: To solve the problem that the MIMLSVM algorithm has low prediction accuracy, a new model was presented to predict the function of protein using an improved MIMLSVM algorithm. Firstly, the modified Hausdorff distance was used to calculate the space distance of each bag. Besides,to improve prediction accuracy, the K-Medoids wad used to covert the MIML (Multi-Instance Multi-Label) question to the multi-label question. Secondly, the multi-label problem was transformed into the multiple independent binary classification problems by using SVM algorithm. According to the characteristics of the protein data, a model was presented to predict the function of protein. Meanwhile, the Particle Swarm Optimization algorithm was referred to optimize the parameters of model. Finally, the experiment was taken on seven organisms. The experimental results showed that the proposed model was superior in the prediction of function of protein.