A learning approach where training data consists of bags (groups) of instances, useful when only bag-level labels are available.