Most one-class classification algorithms measure similarity based on Euclidean distance between samples.Unfortunately, the Euclidean distance couldn' t reveal the internal distribution of some datasets, and so reduced the descriptive ability of these methods. A distance metric learning algorithm was proposed to improve the performance of one-class classifiers in this paper. Compared with existing distance metric learning algorithm, the algorithm only needed to provide target class data, it could effectively solve distance metric learning problem for one-class samples in high-dimensional space by imposing sample distribution prior and sparsity prior with 11-norm constraint on the distance metric,and the formulation could be efficiently optimized in a block coordination descent algorithm.The learned metric can be easily embedded into one-class classifiers, the simulation experimental results show that the learned metric can effectively improve the description performance of one-class classifiers, in particular the description of covering classification model and obtain better generalization ability of one-class classifiers.%已有单类分类算法通常采用欧氏测度描述样本间相似关系,然而欧氏测度有时难以较好地反映一些数据集样本的内在分布结构,为此提出一种用于改善单类分类器描述性能的高维空间单类数据距离测度学习算法,与已有距离测度学习算法相比,该算法只需提供目标类数据,通过引入样本先验分布正则化项和L1范数惩罚的距离测度稀疏性约束,能有效解决高维空间小样本情况下的单类数据距离测度学习问题,并通过采用分块协调下降算法高效的解决距离测度学习的优化问题.学习得到的距离测度能容易地嵌入到单类分类器中,仿真实验结果表明采用学习得到的距离测度能有效改善单类分类器的描述性能,特别能够改善覆盖分类的描述能力,从而使得单类分类器具有更强的推广能力.
展开▼
