基于空间向量分解的边界剥离密度聚类

作为聚类的重要组成部分, 边界点在引导聚类收敛和提升模式识别能力方面起着重要作用, 以BP (Border-peeling clustering)为最新代表的边界剥离聚类借助潜在边界信息来确保簇核心区域的空间隔离, 提高了簇骨架代表性并解决了边界隶属问题. 然而, 现有边界剥离聚类仍存在判别特征不完备、判别模式单一、嵌套迭代等约束. 为此, 提出了基于空间向量分解的边界剥离密度聚类(Density clustering based on the border-peeling using space vector decomposition, CBPVD), 以投影子空间和原始数据空间为基准, 从分布稀疏性(紧密性)和方向偏斜性(对称性)两个视角强化边界的细粒度特征, 进而通过主动边界剥离反向建立簇骨架并指导边界隶属. 与同类算法相比, 40个数据集(人工、UCI、视频图像)上的实验结果以及4个视角的理论分析表明了CBPVD在高维聚类和边界模式识别方面具有良好的综合表现.

Density Clustering Based on the Border-peeling Using Space Vector Decomposition

Border points, as an essential part of density clustering, play a key role in guiding clustering convergence and improving pattern recognition ability. Indeed, the border-peeling clustering with BP (border-peeling clustering) as the latest representative ensures the spatial isolation of core region of the cluster by using intrinsic boundary information, then enhancing the cluster backbone. Nevertheless, the performance of available methods tends to be constrained by incomplete discriminant feature, single pattern and multiple iterations. To this end, this paper proposes a novel algorithm named CBPVD (density clustering based on the border-peeling using space vector decomposition). The property of CBPVD is based on the projection subspace and original space to enhance the fine-grained feature representation of the border point from the two perspectives of sparsity (compactness) and skewness (symmetry) of distribution, then reversely establishes the cluster backbone through active boundary peeling and guides the boundary membership. Finally, we compare performance of CBPVD with six state-of-the-art methods over synthetic, UCI, and image datasets. Experiments on 40 datasets and discussion cases from 4 perspectives demonstrate that our algorithm is feasible and effective in clustering and boundary pattern recognition.