Multi-label And Multi-instance Learning For Improved Classification Algorithms

With the development of modern technologies,such as smart phones,data centers and vast Internet servers,the capability to manually store and process data continues to grow.As a consequence,there has been a tremendous increase in the amount of information stored throughout the world.Therefore,automatic and efficient classifications and tagging of the information become increasingly urgent,which gives rise to great importance of machine learning.Traditional machine learning is based on the assumption that the data are presented with single instance and single label,which is not applicable to the real-world tasks owing to the increasing of the complexity of data structure.In real-world scenarios,it is more meaningful to consider multi label learning(MLL)in which each instance has multiple labels or multi instance learning(MIL)in which multiple instances correspond to single label.In order to facilitate the development of machine learning and to generalize the algorithms and improve their efficiencies,the comprehensive study on MLL and MIL is needed.In this work,we aim to deal with the key problems in the field of MLL and MIL.The main achievements are listed as follows.1.We have taken into consideration the imbalanced distribution in multi-label datasets and proposed an algorithm using correlations among labels to alleviate the imbalance problems.In this model,the multi-label datasets are reconstructed into multiclass ones through coupling with other labels,which is effective to avoid dealing with a tremendous number of potential subsets of the labels individually.Therefore,the relief of imbalance in multi-label datasets is achieved.Meanwhile,to prevent excessive reliance on label correlations,the proposed model assigns a binary classifier for each label and explores its feature.Finally it predicts the multi-labels by integrating the classifiers.The effectiveness of this model,i.e.multi-label learning model based on label correlation and imbalance(MLCI),is sustained both theoretically and experimentally.2.We introduce the transfer learning into MLL and propose a multi-label metric transfer learning model,to further solve the multi-label related problems and improve the performance of the algorithm.This model handles the distribution divergence on both label and instance levels between domains in multi-label datasets,by taking the advantages of transfer learning and employing instance weights to bridge the distributions of training and testing domains.It performs as the foundation of transfer learning in multi-label classifications.3.We add metric into transfer learning to better understand the intrinsic geometric information of label space in multi-label datasets.Therefore,multi-label metric transfer learning(MLMTL)has been put forward.This advanced algorithm can deal with the distribution divergence on both label and instance levels between domains,while retaining the intrinsic geometric information,and consequently improves its performance for prediction.4.We have extended the technological strategy that is adopted in multi-label metric transfer learning,to multi-instance learning scenario.Specifically,to settle the distribution difference on bag level between domains in multi-instance datasets,we propose multi-instance metric transfer learning(MIMTL).MIMTL bridges the distributions between domains by introducing bag weights into multi-instance datasets in training domains.We formulate a new learning principle to adjust the intra-class and inter-class parameters and relax the strict constraint conditions.The model is then built through the utilization of reweighted bags,and the computing efficiency is improved due to the decrease in the number of tunable parameters in common MILs.