Reseach On Graph-based Robust Subspace Learning

In the big data era,the increasing amount of data has stimulated the development of artificial intelligence,especially deep learning.At the same time,it also brings more challenges,e.g.,the processing of high definition images and videos requires more computational power,nosies are mixed in the user-generated data,heterogeneous data are more difficult to be handled and so on.To address these issues,this thesis proposes a battery of subspace learning algorithms based on graphs.As a result,the “curse of dimensionality” can be avoided by mapping the high-dimensional data onto the learning subspace.Meanwhile,the noises mixed in the original data can be filtered out,and the computational costs can be significantly reduced.Specifically,this thesis focuses on the following issues:At first,the purpose and contributions of our research are presented.We briefly review some related work,especially which focus on graph based subspace learning,reported in recent literatures.We discuss the pros and cons of several state-of-the-art approaches.Besides,typical applications,such as image classification,visual tracking,text categorization and multi-modality learning,are discussed.Secondly,we propose a suite of novel graph based subspace learning approaches.We investigate the influence of different graph structures under the framework of graph embedding.Furthermore,we propose several novel graph structures which are applicable for multi-view and multi-modality image classification.Compared with previous methods,our approach not only preserves the inter-class discriminate ability but also cares for the inter-view and inter-modality shared information.Extensive experiments on popular benchmarks verify the superiority of our approach.Thirdly,we investigate the robustness of our algorithm by formulating graph based subspace learning and low-rank representation into a unified optimization problem.As a result,our approach can robustly handle multi-view and multi-modality data,and reveal the true data structure from corrupted samples.Specifically,we exploit the benefits from deploying low-rank constraints on both feature level and samples level.Our methods are evaluated on several standard benchmarks.Fourthly,with the development of mobile Internet,hundreds and thousands of new Apps come online in every single day.For these new Apps,there is a challenging problem of training for the reason that most of practical machine learning algorithms are supervised or semi-supervised.However,there is not enough training samples for new Apps.On the other hand,although deep learning has been widely used in many fields,deep learning needs a lot of training samples.In some specific areas,such as medical image recognition,training samples are limited,which handicaps the applications of artificial intelligence in such areas.To handle the sample insufficient problem,transfer learning has received a lot of attention in recent years.Transfer learning aims at adapting the learned knowledge from related domains to new ones.As a key study of this thesis,we investigate several branches of transfer learning,such as homogeneous domain adaptation,heterogeneous domain adaptation and landmark selection.Fifthly,combining the above technical nuggets,e.g.,subspace learning,low-rank representation and transfer learning,we propose a novel recommendation algorithm which challenges both cold-start recommendation and long-tail recommendation.Our approach learns from social networks,and transfer the learned knowledge to new domain.At the same time,we split the recommended items into a low-rank part to formulate popular items and a sparse part to formulate long-tail items.Experiments on several real-world datasets verify the effectiveness of our approach.At last,we present a brief conclusion of this thesis,and give some inspirations for the future work.