“…To obtain a good performance in the test process, the problem of learning a classifier to directly optimize the same multivariate performance measure over the training set is proposed [1,2,3,4,5,6,7,8,9,10,11,12,13]. Meanwhile, the key of representing multi-instance data is the learning of a multiinstance dictionary, which is used to map a bag to a bag-level feature vector [14,15,16,17,18,19,20,16,21,22,23,24]. We observe that different performance measures require different optimal multi-instance dictionaries, i.e., the optimal performance is dependent on the optimal dictionary.…”