Research On Incomplete Modality Transfer Learning Methods

Transfer learning has been proven to be an extremely effective technique in many realistic application scenarios such as image classification,text classification,sentiment analysis and so on.However,most of this is based on the assumption that the training dataset is complete.When the training dataset has missing data or data corruption,traditional transfer learning methods are stretched to deal with such problems.If the model training is forced,it may lead to bad classification results.In recent years,the field of transfer learning has developed rapidly.Transfer learning in incomplete data scenarios has attracted the attention of some researchers.They have explored the combination of subspace learning framework and latent low-rank constraints to deal with this scenario,and achieved good results.Due to the complexity and changeability of realistic scenarios,such as light,angle,shadow and so on,which will negatively affect the data,it is a great challenges for data collection.The solution to train transfer learning model with datasets in extreme scenarios is a problem worthy of attention.Traditional transfer learning algorithms are powerless in this challenging scenario.In order to solve the above problems effectively,subspace learning through latent low-rank constraints framework was proposed for incomplete modality scenario and incomplete multiview scenario.This thesis can be devided as follows:(1)The transfer learning algorithm called IMTL(Incomplete Modality Transfer Learning via Latent Low-rank constraint)is proposed in this study.The contributions are twofold:(1)After latent factors are introduced into a low-rank constrained subspace framework so as to mine missing modality information on the target domain,with the help of an auxiliary yet complete modality dataset,the proposed cross-modality and cross-dataset transfer learning strategy is used to help align data between modalities or datasets.(2)A small amount of labeled target data is used to align the supervision information so as to maintain the internal structure of the target data during the transfer learning.Experimental results show that the proposed transfer learning algorithm has obvious advantages over traditional transfer learning algorithms on the adopted incomplete target datasets.(2)The subspace learning through latent low-rank constraints framework was extended from incomplete modality problem to incomplete multi-view problem,and the multi-view learning algorithm called IMSL(Incomplete Multi-View Subspace Learning through Dual Low-Rank Decompositions)is proposed in this study.The contributions are twofold:(1)Latent factors are introduced into dual low-rank decompositions subspace framework so as to mine missing information in the multi-view data.(2)IMSL aims to seek a more robust subspace through pre-learned low-dimensional features of multi-view data.Furthermore,the supervised information is used to guide dual low-rank decompositions.Experimental results show that the proposed algorithm has obvious advantages over previous multi-view subspace learning algorithms on the adopted incomplete multi-view datasets.