Research On Two Algorithms For Multi-label Feature Selection

Different from traditional supervised learning framework in which each instance is associated with one label,multi-label learning framework in which one instance may be associated with multiple labels simultaneously is able to analyze the problems in real world more effectively.Multi-label learning which derived from the problem of multi-label text categorization is a challenging research topic.For nearly a decade,multi-label learning has attracted the interest of a growing number of researchers and has been widely applied to diverse problems such as bioinformatics,automatic annotation for multimedia contents and sentiment classification etc.Researchers have proposed a series of multi-label learning algorithms.Nevertheless,in high dimensional data,there are a lot of irrelevant features and redundant features which reduce the performance of classifiers.So feature selection,a highly efficient technique for dimensionality reduction,plays a core essential role among multi-label classification.However,there are few researches for the problem of multi-label feature selection nowadays.In this respect,this dissertation focuses on this problem and proposed two multi-label feature selection algorithms by employing non-negative matrix factorization and manifold learning techniques to multil-label learning.The main contributions of this dissertation are as follows:1.We propose a multi-label feature selection algorithm based on non-negative sparse representation.First of all,we employ subspace learning to realize multi-label feature selection,and impose the indicator matrix with non-negative and sparsity constraints in the process of matrix factorization.Then,a kind of efficient iterative update algorithm is designed to tackle this fusion optimization problem which is combined with non-negative constraint and 2,1L-norm minimization constraint.Finally,experimental results demonstrate that proposed algorithm has better performance of feature selection on multi-label data.2.We propose a multi-label feature selection based on feature manifold and sparse regularization.Firstly,with the least square regression model,we formalize multi-label feature selection as a matrix factorization problem with regularized term using regression coefficient matrix to assess the importance of different features.Secondly,we embed feature manifold and sparsity constraint into optimization problem to capture smooth and sparse regression coefficient matrix.Finally,an efficient iterative update algorithm is designed to tackle this problem,and its effectiveness is shown by experimental results.