论文部分内容阅读
目标数据呈簇分布、基于超平面的单类分类器要求嵌入结构信息时,必须分别考虑各簇数据对超平面的影响,为此,提出可用于簇分布的结构大间隔单类分类器(structural large margin one-class classifier,SLMOCC)。该算法通过分别约束各簇数据到超平面的马氏距离,并最大化最小马氏间隔,保证目标数据落入正半空间的同时,充分利用数据的簇结构信息,通过序列二次锥规划优化方法线性搜索到最优超平面。为捕捉数据簇结构,SLMOCC采用凝聚型层次聚类并借助拐点确定聚类数目,最后通过人工数据和UCI数据集与相关算法比较,验证了SLMOCC的有效性。
When the target data is distributed in clusters, when the single-plane classifier based on hyperplane needs to embed the structure information, the impact of each cluster data on the hyperplane must be considered separately. For this reason, the structural large-interval single classifier large margin one-class classifier, SLMOCC). By constraining the Mahalanobis distance of each cluster data to the hyperplane and maximizing the minimum Markov space, the algorithm ensures that the target data fall into the positive and negative half-space while making full use of cluster structure information of the data, The method searches the optimal hyperplane linearly. In order to capture the data cluster structure, SLMOCC uses condensed hierarchical clustering and determines the number of clusters by means of inflection points. Finally, the validity of SLMOCC is verified by comparing artificial data and UCI data sets with related algorithms.