Research On Feature Selection Algorithms Based On Structure Information Of Samples And Features

As a key dimensionality reduction technique in pattern recognition, feature selectionhas beenwidelyused in information retrieval, text classification and genedata analysis. In recent years, usingstructural information contained in samples to guide feature selection has become a new hot spot infeature selection study, such as Laplacian Score method. In this thesis, we first take into account thestructural information of samples and propose two iterative feature selection algorithms based oncertain structure of samples. The former uses the local structure of samples, while the latter uses thesparse reconstruct relationship of samples. Additionally, the choice of how to take advantageofstructural relationship between data features as guidance in feature selection is another focus of thisthesis. We propose a group Lassoalgorithm based on feature clustering. The main innovation andresearch work are summarized as follows:Firstly, we propose an iterative Laplacian score algorithm based on the local structure of samples.In each iteration process, it evaluates the features’ ability of locality preserving, and discards the mostunimportant features, then reconstruct the model of local structure to achieve the purpose ofoptimizing feature subset.Secondly, taking advantage of the sparse reconstruct relationship of samples and coupled withiteration, we propose an iterative sparsity score algorithm based on the sparse reconstruct relationshipof samples. It continuously optimizes the model of sparse reconstruct relationship of samples in aniterative manner, in order to achieve enhancement of the performance.Thirdly, we apply K-means clustering algorithminto feature selection and propose a groupLassoalgorithmbased on K-means clustering for features. This algorithm discovers the structuralrelationship between the features by clustering, and then combined with group Lasso algorithm toenforce structural feature selection.