Research Of View-Learning Based Classification Methods

In machine learning classification methods,the data is represented by a set of features which can be obtained from single-view or multi-view.According to the number of views used to describe problems,we can classify the machine learning methods into two groups:single-view learning methods and multi-view learning methods.In many application sce-narios,data is always described by single-view features,and the single-view data can be seen everywhere.With the development of technology,data can be collected from diverse domains or obtained from various feature extractors.The ways to describe a problem are becoming more and more diverse and we tend to have more and more multi-view data.In real-world applications,we need to analyze many different single-view datasets and multi-view datasets.Meanwhile,in the era of big data,with the rapid development of data ac-quisition equipments,we have more and more high-scale or high-dimensional multi-view data.For dealing with more and more complex multi-view data,many traditional multi-view learning methods can not achieve an satisfactory result as the high computational cost or be lack of ways to deal with high-dimensional data.To process the single-view and multi-view data in real applications,we have three ways,namely using single-view learn-ing methods to process single-view data,using single-view learning methods to process multi-view data and using multi-view learning methods to process multi-view data.Based on the three processing ways,this thesis,from the perspective of view learning,studies and exploits the single-view classification methods and multi-view classification methods by proposing several novel algorithms.And we exploit how to extend single-view learning to multi-view learning scenarios.For processing single-view data with single-view learning methods,we introduce two genetic algorithm based single-view classification methods which improve models' gen-eralization performance and robustness.In the first method,we use GA to optimize the randomly generated parameters of ELM and then use a sorting method designed by the generalization theory of ELM to select a set of good ELMs to make ensemble.The ex-perimental results reveal that GE-ELM can improve the robustness and the generalization performance of ELM.In the second method,we utilize GA's global search ability and take the advantages of ELM to propose a novel feature selection method(HGEFS,Hybrid ge-netic algorithm and extreme learning machine for feature selection).HGEFS improves the performance of feature selection through improving the efficiency of search strategy,opti-mizing the structures of networks for each feature subset and making full use of the statis-tical information for each feature.For processing multi-view data with single-view learning methods and to better pro-cess large-scale multi-view data,we exploit a way to integrate single-view learning meth-ods and multi-view learning methods by proposing a single-view based multi-view learning method called LMVL(a linear computational cost multi-view learning method).LMVL op-timizes the objective function to convert the multi-view learning problem to a set of linear single-view learning optimizing problems.LMVL is a linear computational cost and par-allelized multi-view learning method.In LMVL,we simultaneously analyze all features by learning an integrated projection matrix and we can automatically select more impor-tant views for predicting different classes.Compared to the conventional methods which learn the entire projection matrix,our algorithm independently optimizes each column of the projection matrix for each class,which can be easily parallelized.As a parallelized with a linear computational cost,LMVL is quite applicable to large-scale problems.To better process high-dimensional multi-view data,we propose a novel subspace-learning based multi-view method called MVSC(Multi-view feature learning with shared component).Compared to most of the existing subspace learning methods that only focus on exploiting a shared latent subspace,our algorithm not only learns individual information in each view but also captures feature correlations among multiple views by learning a shared component.In this case,MVSC can more effectively utilize the rich information of different views.In MVSC,we use different projection matrixes to project the shared information and individual information into different subspace.Exploiting the useful information in these subspaces can effectively reduce the dimensions of each view.In this case,MVSC is applicable to high-dimensional multi-view datasets.Since the objective function is non-smooth and difficult to solve,we propose an efficient iterative algorithm for optimization with guaranteed convergence.Extensive experiments are conducted on several benchmark datasets.The results demonstrate that MVSC performs better than all the compared multi-view learning algorithms.