Research And Application Of Multi Instance Multi Label Active Learning Method

Active learning is an effective approach to solve the problem of label missing in traditional classification problems.It has got much attention in scientific studies or practical applications areas.Multi instance multi label(MIML)learning is a state-of-the-art machine learning framework,which is quite qualified for complex tasks in real world.However,researches about MIML active learning is still a blank.This thesis aims to carry on some researches and applications of multi instance multi label active learning method.The main work includes:(1)MIML active learning framework based on label ranking(MIMLAL)is proposed.Considering the traditional active learning frameworks,the select strategies and the MIML framework,MIMLAL is designed to reduce information loss and the compute in MIML active learning problems.(2)Two different MIML active learning methods are proposed: the MIMLAL-A method based on the max mean discrepancy in label space and the MIMLAL-R method based the rank of instance-level predictions.In MIMLAL-A,a new select strategy named MILCI based on the max mean discrepancy(MMD)between unlabeled and labeled samples is proposed.It is extended from label cardinality inconsistency(LCI)to give a more accurate prediction of unlabeled sample bags rather than the traditional strategy based on key instance.To go further,a strategy based on the rank of instance-level predictions named RLCI is proposed in MIMLAL-R to avoid the irrelevant information taken in by MILCI.To enhanced the ability to distinguish the values of different sample bags,RLCI is combined with the label-selecting strategy to select bag-label pairs in MIMLAL-R.(3)The MIMLAL-A and MIMLAL-R methods are implemented to predict protein functions.First,the task of predicting protein functions is abstracted as a MIML learning problem.Then,the methods MIMLAL-A and MIMLAL-R are used to solve this problem.Their performances are compared with some multi label active learning methods,and the result indicates that the MIML active learning methods proposed in this thesis get better performance,especially the MIMLAL-R method.