Research On Multi-View Learning Approaches Towards Data Privacy Preserved

With the rapid development of the Internet,a huge number of rich media objects can be collected from various information channels or described by multi-view feature groups.Multi-view learning,which attempts to exploit the relationship among multiple views to improve learning performance,has drawn extensive attention.It is noteworthy that in some real-world applications,features of different views may come from different private data repositories,and thus,it is desired to exploit view relationship with data privacy preserved simultaneously,i.e.not to share data features with other data repositories when training and testing the model.At the same time,this cooperation-competition scenario raise a demand of enhancing the performance of view-specific classifier,in addition to a good ensemble result.Towards data privacy preserved during multi-view learning,there are two main contributions in this paper:1.In order to both improve the performance of multi-view learner and view-specific learner,we propose a novel multi-view learning framework which works in a hybrid fusion manner based on the consistency of the predictions of multiple views.An improved fused result can be learned by converting predicted values of each view into an Accumulated Prediction Matrix(APM)with low-rank constraint,which is delivered back to update parameters of view-specific model.The learning phase is carried on predictions of each view without accessing features of these views,so we provide a way to preserve data privacy in multi-view learning.Furthermore we consider variants of solutions for the optimization and an acceleration implementation.Experimental results shows that our model achieves significant better performance and high efficiency on various tasks.2.Making a better use of unlabeled data can rely less on labeled data from private information channel.Under the framework of co-training,by using view consistency defined on all the other view-specific classifiers' predictions on new unlabeled examples as the confident score for being selected to enlarge training set of one view,this paper propose a novel view consistency criterion for view-specific learners to improve one another.The access between views are limited in the training phase,and the requirement of ability of initial classifiers can be further reduced with the involvement of much unlabeled data.Experimental results verify the effectiveness of the proposed method on most datasets.