A Study On Multi-label Transfer Learning Algorithm And Application In The Early Diagnosis Of Alzheimer’s Disease

An important prerequisite for the success of multi-label learning is that there is a certain amount of multi-label training samples.When the samples are scarce,the model is prone to overfitting,so multi-label transfer learning is born.Multi-label transfer learning aims to use multi-label samples in similar domains to assist in training the target domain to train the classifier.Multi-label feature transfer is an important branch in multi-label transfer learning.Current research mainly focuses on how to narrow the differences between domains,usually in the feature subspace to narrow the probability distribution differences of cross-domain data,so as to learn common feature representations.Existing multi-label feature transfer algorithms always violently mine the links of all labels.Although in theory,the correlation between different labels is fully utilized to fit the real model,but when there are too many labels,the model faces training.The number of categories will increase explosively.For a certain number of small-scale training samples,the number of training samples corresponding to each category is further reduced,and the problem of uneven distribution of sample data between classes is highlighted,which greatly increases the difficulty of classifier training.In addition,the existing multi-label feature transfer algorithm usually uses the narrowed edge probability distribution between domains to map the common feature subspace,but only narrowing the edge distribution cannot meet the application requirements of all scenarios.In particular,when the overall domains are similar but the categories are not similar Although the distribution difference between the domains is small,the large difference between the categories still makes the degree of matching between the domains vary greatly during transfer.Existing algorithms cannot be applied to this scenario.This study will study the above technical issues.At the same time,in the early diagnosis of Alzheimer’s disease(AD)based on multi-modality,samples with marked information are scarce,so multi-marker feature transfer often transfers the disease characteristics of the advanced stage to the early diagnosis.According to research,the thickness of the cortex reflects the atrophy of different brain regions and has a strong discriminating ability.However,not all the cortical thickness features of the brain area contribute to the early AD diagnosis model,that is,the introduction of cortical thickness will bring certain redundant features.Existing algorithms cannot solve the problem of redundancy during feature transfer.This paper improves the previous algorithm And apply it to early diagnosis of AD.The contributions of this paper are as follows:(1)Aiming at the problem of poor generalization ability of existing multi-label feature transfer,propose a multi-label transfer learning algorithm via joint distribution(L-MLTL).In order to solve the learning difficulties caused by the large label space,a multi-label transfer learning method based on label space decomposition(Multi-label Transfer Learning Via Label Label Space Decomposition,L-MLTL)is designed,using label consistency matrix decomposition technology(Label Consistent Matrix Factorization(LCMF)decomposes the original label space into a complete label matrix and a coding matrix,and uses the coding matrix to replace the original label matrix for multi-label feature transfer.Through transfer experiments on the benchmark datasets Corel5K,ESPGame and Iaprtcl2,L-MLTL’s Hamming loss,Ranking loss,Coverage and Average Precision indicators are better than the comparison algorithm by about 4.2%,9.8%,2.6%and 1.1%,notable Yes,when the number of labels is greater than 4,the classification performance of L-MLTL will be better than existing algorithms.The results show that the coding matrix has better discriminative ability than the original labeling matrix,which will reduce the learning complexity of the model.(2)In order to solve the problem of poor generalization ability of existing multi-label feature transfer,,propose a multi-label transfer learning algorithm via joint distribution(J-MLTL).When the overall domains are similar but the categories are not similar,the algorithm Performance is limited.Therefore,this paper designs a multi-label transfer learning via joint distribution alignment(J-MLTL).By reducing the marginal distribution and conditional distribution of cross-domain data to reduce the difference between domains,at the same time using hypergraph learning to strengthen the correlation between multiple markers,so as to ensure a good transfer effect.Through transfer experiments on the benchmark datasets Corel5K,ESPGame,and Iaprtc12,compared with S-MLTL,M-MLTL and other algorithms,J-MLTL’s Hamming loss,Ranking loss,One-Error,Coverage and Average Precision indicators have increased by 5.7%、11.1%、2.8%、21.6%and 7.2%respectively.Experiments show that the measurement method based on joint distribution is significantly better than a single marginal distribution or conditional distribution,effectively improving the classification accuracy.(3)In the early diagnosis of Alzheimer’s disease with a small sample,in order to solve the problem of feature redundancy caused by cortical thickness,propose an early AD diagnosis based on two-stage transfer learning(2STL).In order to find the most discriminating cerebral region cortical thickness from the redundant cortical thickness feature,and further improve the model learning efficiency and classification accuracy,an early diagnosis of AD based on two-stage transfer method(Early diagnosis of AD based on two-stage transfer learning,2STL).In this study,the minimum redundant maximum correlation(mRMR)method is introduced in the feature subspace of J-MLTL to eliminate redundant features,in order to find the most discriminant among the redundant cortical thickness features The cortical thickness of the brain area further improves the model learning efficiency and classification accuracy.Finally,the TrAdaBoost algorithm is used for instance transfer to reduce the impact of irrelevant samples on classification.For the ADNI dataset,the classification accuracy of 2STL is 68.6%,which is better than M2LTL and rMLTFL by about 2.1%.Throughexperimental verification,the removal of redundant features can improve the accuracy of transfer,and the performance of two-stage transfer(feature transfer+instance transfer)is better.