Research On One-class Classifier Based On Geometric Covering Model Of Target Class In Low-dimensional Subspace

The essence of one-class classification is to distinguish the target class from all theother classes using training data from the target class. Up to now, the research of one-classclassification has attracted much attention. However, due to redundant and noisy inhigh-dimensional data, a covering model constructed from these data can not reflect theirdistributing information, which lead to the classification performance of irregular andcomplex data in high-dimensional space is descended. Therefore， based on theachievements of previous researchers, this paper researches on conducting some novelcovering models of one-class classifier in low-dimensional subspace.Firstly, an ensemble multi-trees algorithm in high-dimensional space based pruningrandom subspace method is proposed to improve the descriptive performance of one-classclassifiers in this paper. Firstly several random subspaces are created, and constructedminimum spanning tree covering models independently in each subspace. Next, it appliespruning ensembles to each classifier by using an evaluation criterion. And finally, itintegrates these subspace classifiers as an ensemble classifier by mean combining.Experimental results show good performance of this method.Secondly, in order to construct the tight covering model of target class, an approxi-mate convex hull covering model based dimensionality reduction by sparse representationis proposed. Firstly the homotopy algorithm is used to solvel1norm problem, neighborsare automatically captured based sparse constraint then neighborhood graph is constructed.Next, LPP is applied in order to fast and efficient dimensionality reduction. And finally, anapproximate convex hull covering model is constructed in low-dimensional space andrealized one-class classification. Experimental results demonstrate the effectiveness of thismethod.Finally, a manifold minimum spanning tree covering model is proposed in this paper.Firstly neighbors are automatically captured based sparse representation, and thenneighborhood graph is constructed. Next, sparsity preserving projection is applied in orderto fast and efficient dimensionality reduction. And finally, minimum spanning tree covering model is constructed in low-dimensional space and realized one-classclassification. Experimental results show good performance of this method.