Abstract:  In dealing with multiple semantic objects，multi－instance multi－label learning overcomes the shortcomings of the multi－instance learning and multi－label learning，which are applied to text classification，image recognition，analysis of genetic data and so on. There are two main degeneracy strategies based on the stage of degeneration. One of the degeneration strategies is to use the multi－label learning algorithm as a bridge. By using K－Medoids clustering algorithm，the multi－instance multi－label problem is transformed into a single－instance multi－label problem. And some state－of－the－art effects have been achieved through this degenerate strategy approach. At the same time，this approach simplifies the handling of multi－semantic and complex semantic objects. In this degradation strategy process，some important data information of multi－instance multi－label is weakened or even lost，which is due to a lack of consideration of the correlation between each instance. In order to overcome this drawback and replace K－Medoids，an improved algorithm is proposed for multi－instance multi－label learning based on mean shift clustering(MIMLMS). This MIMLMS algorithm is a degenerate method using the mean shift algorithm，which adds Gaussian kernel functions and weights. Therefore，the dependency relationship between each instance is obtained by adding instances weights. And the multi－instance bags are introduced into the Gaussian kernel function，which can use the kernel density estimation and gradient descent method to solve the optimal solution of the degradation strategy process. Finally，the multi－instance multi－label classification model is established by using the objective function of the sum of squared errors(SSE). The experimental results of the MIMLMS algorithm are given in the benchmark multi－instance multi－label test data set，which fully show that the algorithm has good classification performance，effectiveness and reliability.