Research On One-class Classification Algorithm Based On Projection Subspace

One-class classification problem such as identification and anomaly detection, isdifferent from traditional pattern recognition. These problems only use the commonalityinformation among the same mode to build an effective covering model to completeclassification tasks. However, due to redundant and noisy in high-dimensional data, acovering model constructed from these data can not reflect their distributing information,which lead to the classification performance of irregular and complex data inhigh-dimensional space is descended. Therefore, based on the achievements of previousresearchers, this paper researches on how to obtain the low dimension representation,explore the inherent law and nature structure of high-dimensional data and thenconducting some novel covering models of one-class classifier in low-dimensionalsubspace.Firstly, spherical cover classification algorithm based on manifold dimensionreduction space of local and global regressive mapping is proposed in this paper. Themapping model which combines local information with global information is extractedfirstly, and the local laplacian matrix and global laplacian matrix are optimized separately,to obtain low dimension representation of train data by eigen-decomposition of laplacianmatrix, then the low dimension representation of test data with kernel mapping is obtainednaturally. The spherical cover classification model in low dimension space is constructedin the last step. Experimental results demonstrate the effectiveness and feasibility of themethod.Secondly, NN cover one-class classification algorithm based on kernel sparsitypreserving projections is proposed in this paper. The algorithm obtains kernel sparserepresentation coefficient in the high dimensional space by kernel method and use kernelsparse representation coefficient which contains strong discriminant information toconstruct adjacency matrix. To obtain the low dimension representation of original data,an adjacency is necessary. The NN cover classification model in low dimension space isconstructed in the last step. Experimental results show good performance of this method.Finally, NN cover classification algorithm based on linear projection of iterative nearest neighbors is proposed in this paper. The algorithm treats each training sample ascurrent query, and use the reconstruction of current query which is calculated by iterationin the rest training points to construct adjacency matrix. To obtain the low dimensionrepresentation of original data, an adjacency is necessary. The NN cover one-classclassification model in low dimension space is constructed in the last step. Experimentalresults show the feasibility and effectiveness of this method.