Research On Feature Missing Problem Based On Low-Rank Learning

Data analysis,which extracts useful information and knowledge from massive data,has been developing rapidly.Now it has been widely used in all walks of life.Traditional data analysis methods are designed to manipulate complete data sets.However,in practical applications,due to environmental disturbance or level of technic,the missing of data features is often inevitable,which makes it difficult for traditional data analysis and processing methods to learn accurate and complete learning models from such data.How to directly learn from the data with feature missing becomes an important research problem in the fields of data mining,artificial intelligence and machine learning.Based on low-rank learning,this paper studies the problem of data feature missing in multi-view and multi-instance learning,which both are the typical scenario where feature missing often occurs.The main work is as follows:(1)Multi-view clustering with feature missing data based on Low-rank learningFor multi-view data with feature missing of view,we propose a structured graph based clustering method based on low-rank learning.It combines incomplete multi-view data and clusters it simultaneously by learning the ideal structures.More specifically,for each view,with an initial input graph,it excavates a clustering structure with the consideration of consistency with the other views.The learned structured graphs have exactly c(the predefined number of clusters)connected components so that the clustering results can be obtained without requiring any post-clustering.An efficient optimization strategy is proposed which can solve both global and local regularization problems.The method is more effective than the traditional method in the contrast experiment.The parameter experiment verifies the robustness of the method and the convergence experiment verifies the convergence of the method.(2)Multi-instance classification with feature missing data based on Low-rank learningFor the new problem of data feature missing in Multi-Instance Learning(MIL),we propose a low-rank learning based framework,Fragmentary multi-Instance Classification(FIC).It jointly completes the fragmentary data and learns the multi-instance classifier.More specifically,we propose the weight mechanism to measure the importance levels of different instances,which effectively facilitates the integration between completion and classifier learning.Three typical MIL methods(MI-SVM,EM-DD,Citation-KNN)are embedded into the framework obtaining three FIC models,to validate the compatibility of our framework.In addition,we propose the different weight functions for different models based on the meaning of ”more positive” in different MIL methods.An effective algorithm is proposed and its convergence is proved.Comparison experiments and parameter study verify the validity and robustness of the framework.