Research On Unsupervised Multi-view Feature Selection Methods

Unsupervised feature selection is an important research topic in data mining and machine learning.It has been widely used in pattern recognition,signal processing,image processing and multimedia retrieval.Unsupervised feature selection for multi-view data can not only filter out noise and redundant information effectively,but also select features with important semantic and structural information in the absence of label information.It is beneficial to the subsequent task,such as clustering,classification,etc.Among the existing unsupervised multi-view feature selection methods,the graph-based methods have achieved satisfactory performance.However,they still suffer from several important problems:(1)The existing graph-based algorithms often directly use the original data to generate a fixed similarity graph.Since the original multi-view data is often mixed with noise and outliers,the initialized similarity graph may bring unreliable semantic structure information and result in undesired effect;(2)The similarity graph learning and the feature selection are two separate and independent steps.Without the interaction,it is often impossible to achieve the optimum of them at the same time;(3)Since multi-view data is characterized by multiple features and it is high-dimensional,existing methods generally have the problem of low computing efficiency,which imposes considerable obstacles to the practical application of feature selection.In order to solve the above three problems,two unsupervised feature selection methods are proposed based on the multi-view data:(1)An effective unsupervised multi-view feature selection method is proposed based on the nonnegative structured graph learning.The method can learn structured graph,nonnegative pseudo labels and feature selection matrix simultaneously.Firstly,it uses adaptive weights to learn the shared structured graph of the original data.Secondly,the nonnegative pseudo labels which involve rich discriminative semantic information is learned with the structured graph guidance.Finally,the nonnegative pseudo labels can supervise the feature selection process in unsupervised scenarios.In addition,by imposing reasonable rank constraint on the Laplacian matrix of the similarity graph,the number of connected components in the learned similarity graph is equal to the number of the data classes exactly.Extensive experiments on several widely tested benchmarks demonstrate the superiority of the proposed method compared with several state-of-the-art approaches.(2)An effective unsupervised multi-view feature selection model based on shared-label learning is proposed.The model integrates shared-label learning and feature selection process into a unified learning framework.In this model,the shared cluster center and the nonnegative shared-label are learned simultaneously to achieve the optimal value.Then the learned nonnegative shared-label can provide necessary supervision information for feature selection task in the absence of label information.The shared-label learned by this method indicate the affinity degree of each sample point for all clusters.And it also solves the problem of low operating efficiency in existing graph-based methods.Experiments on multiple multi-view datasets show that the new model achieves superior performance in terms of operating efficiency and clustering accuracy.