[信息与通信]105-模式识别-聚类算法-05.pptVIP

OTHER CLUSTERING ALGORITHMS The following types of algorithms will be considered: Graph theory based clustering algorithms. Competitive learning algorithms. Valley seeking clustering algorithms. Cost optimization clustering algorithms based on: Branch and bound approach. Simulated annealing methodology. Deterministic annealing. Genetic algorithms. Density-based clustering algorithms. Clustering algorithms for high dimensional data sets. GRAPH THEORY BASED CLUSTERING ALGORITHMS In principle, such algorithms are capable of detecting clusters of various shapes, at least when they are well separated. In the sequel we discuss algorithms that are based on: The Minimum Spanning Tree (MST). Regions of influence. Directed trees. Minimum Spanning Tree (MST) algorithms Preliminaries: Let G be the complete graph, each node of which corresponds to a point of the data set X. e=(xi,xj) denote an edge of G connecting xi and xj. we?d(xi, xj) denote the weight of the edge e. Definitions: Two edges e1 and e2 are k steps away from each other if the minimum path that connects a vertex of e1 and a vertex of e2 contains k-1 edges. A Spanning Tree of G is a connected graph that: Contains all the vertices of the graph. Has no loops. The weight of a Spanning Tree is the sum of weights of its edges. A Minimum Spanning Tree (MST) of G is a spanning tree with minimum weight (when all we’s are different from each other, the MST is unique). Minimum Spanning Tree (MST) algorithms (cont) Sketch of the algorithm: Determine the MST of G. Remove the edges that are “unusually” large compared with their neighboring edges (inconsistent edges). Identify as clusters the connected components of the MST, after the removal of the inconsistent edges. Identification of inconsistent edges. For a given edge e of the MST of G: Consider all the edges that lie k steps away from it. Determine the mean me and the standard deviation σe of their weights. If we lies more than q (