聚类分析与生物分子动力学.pptVIP

聚类分析与生物分子动力学I 姚 远 高等统计选讲 2009.9.18 摘 要 分子动力学仿真数据特点 聚类分析 合作伙伴 Stanford University Guibas-Carlsson-Pande Collaboration 生物分子动力系统仿真 Shirts and Pande, Science, 290, 1903 , 2000 分子动力系统中的聚类分析 Geometric Clustering (距离度量) K-means/K-medoids vs. K-center, etc. Kinetic Clustering Spectral clustering, etc. 聚类分析的性质 Flat clustering vs. Hierarchical clustering Batch vs.Streaming (online) data 近似算法和计算复杂性 统计性质 几何聚类分析GEOMETRIC CLUSTERING Two Types of Clustering How is the data presented Online/streaming n or endless data point, one at once (o(1) or o(n) memory, can NOT store all of them) K-center vs. K-means A Greedy Algorithm for K-center K-center 几何性质 K-center形成了样本空间的一个epsilon-net Any two points in C are R-distance away Points in C form a R-cover of sample space 只依赖于度量结构 K-center is NP-hard, but greedy algorithm is O(kn) K-center在ISOMAP(TdL’2000, Science)中被采用，称为Landmark技术 Molecular dynamics application [Sun, Y, Huang, et al. JPC, 09] 缺点： 对样本空间边缘的outlier和noise比较敏感 Approximability of K-center A Greedy Algorithm for K-means A Constant-factor Approximation Complexity of K-means Hierarchical Clustering A Basic Existence Problem Hierarchical K-center Hierarchical K-center: Complexity Hierarchical Clustering: Open Problems Clustering online/streaming data Online K-center Online K-center Implementation Cover Tree Streaming K-means: I Streaming K-means: I (bicriterion) Streaming K-means: II Hierarchical Agglomerative Clustering Guarantees for Agglomerative Clustering Complete linkage has underlying k-center cost function. Approximability characterization: for all k, the induced k-clustering is within factor ?(k) of the optimal k-center solution… what is ?(k) ? Statistical Theory for Clustering I. Consistency of K-means Consistency of K-means Suppose data Dn={Xi: I=1,…,n} is drawn iid from an underlying distribution P. Let Ck be the optimal k-means centers with respect to P. Let Cnk be the optimal k-means centers for Dn. Density Cluster Tree For any density p(x), consider the s