Abstract:Multi-Instance Multi-Label (MIML) learning is a novel machine learning framework in which an instance is described by multiple instances and associated with multiple labels. This framework has become a hot topic in the field of machine learning because of its excellent expressive ability for polysemous objects. The most direct way to solve the MIML classification problem is the degradation strategy, it takes the multiple instance learning or multiple label learning as a bridge, transforms the MIML problem into a series of binary classification problems. However, the correlation information among labels will be lost in the degradation process, which will affect the classification result. Based on these problems, this study proposes the MIMLSVM-LOC algorithm. The algorithm combines the improved MIMLSVM algorithm with a local label correlation method ML-LOC which considers the correlation information among labels in the training process. The algorithm first improves the K-medoids clustering algorithm in the MIMLSVM algorithm, and then uses the mixed Hausdorff distance to transform each instance packet into an instance, which degradate the MIML problem. Then, the ML-LOC algorithm is used to continue the classification work. In the experiment, the comparison experiment with other MIML algorithms, the result shows that the improved algorithm has better performance than other classification algorithms.