Scribd Logo
Upload

Welcome to Scribd!

  • 34 views

    Kruskal and Clustering

    Uploaded by

    Hadi Fouani
    Kruskal's algorithm can be used to solve the clustering problem by finding clusters with maximum spacing between objects. It works by growing a graph connecting the closest objects, representing emerging clusters. As the graph grows, only connections within existing clusters are added. When the graph has k connected components, k clusters have been formed with maximum spacing equal to the weight of the next connection Kruskal's would have added. This clustering method is equivalent to computing the minimum spanning tree and removing the k-1 most expensive edges to form k clusters.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    Kruskal and Clustering

    Uploaded by

    Hadi Fouani
    34 views1 page
    Kruskal's algorithm can be used to solve the clustering problem by finding clusters with maximum spacing between objects. It works by growing a graph connecting the closest objects, representing emerging clusters. As the graph grows, only connections within existing clusters are added. When the graph has k connected components, k clusters have been formed with maximum spacing equal to the weight of the next connection Kruskal's would have added. This clustering method is equivalent to computing the minimum spanning tree and removing the k-1 most expensive edges to form k clusters.

    Original Description:

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Kruskal's algorithm can be used to solve the clustering problem by finding clusters with maximum spacing between objects. It works by growing a graph connecting the closest objects, representing emerging clusters. As the graph grows, only connections within existing clusters are added. When the graph has k connected components, k clusters have been formed with maximum spacing equal to the weight of the next connection Kruskal's would have added. This clustering method is equivalent to computing the minimum spanning tree and removing the k-1 most expensive edges to form k clusters.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    34 views1 page

    Kruskal and Clustering

    Uploaded by

    Hadi Fouani
    Kruskal's algorithm can be used to solve the clustering problem by finding clusters with maximum spacing between objects. It works by growing a graph connecting the closest objects, representing emerging clusters. As the graph grows, only connections within existing clusters are added. When the graph has k connected components, k clusters have been formed with maximum spacing equal to the weight of the next connection Kruskal's would have added. This clustering method is equivalent to computing the minimum spanning tree and removing the k-1 most expensive edges to form k clusters.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 1

    Kruskal’s Algorithm and Clustering

    (following Kleinberg and Tardos, Algorithm design, pp 158–161)

    Recall that Kruskal’s algorithm for a graph with weighted links gives a minimal span-
    ning tree, i.e., with minimum total weight. This solves, for example, the problem of
    constructing the lowest cost network connecting a set of sites, where the weight on the link
    represents the cost.
    The minimal spanning tree is also relevant to the clustering problem. Given some
    notion of similarity between objects, it is frequently useful to group objects into clusters,
    where clusters contain objects that are in some sense most similar. The objects could be
    photographs, documents, micro-organisms, . . .. Given a set of objects p1 , . . . , pn , a distance
    function d(pi , pj ) specifies their similarity (or lack thereof). This function is symmetric:
    d(pi , pj ) = d(pj , pi ), and satisfies d(pi , pj ) ≥ 0 (=0 iff i = j).
    Suppose the n objects are to be separated into k clusters C1 , . . . Ck . The “spacing” of
    any particular clustering is defined as the minimum distance between objects in any pair of
    different clusters. One reasonable criterion for a “good” clustering is to find k clusters with
    maximum spacing. Since the number of possible clusterings grows exponentially with the
    number of objects, we need an efficient algorithm to find the one with maximum spacing.
    Consider growing a graph on the objects pi considered as vertices. Start by drawing
    an edge between the closest pair of points, then next closest, etc., and at any given con-
    figuration the sets of connected vertices represent the clusters. (This procedure is known
    as single-link agglomerative clustering.) Note that since only the sets of clusters are of
    interest, it is not necessary to add any edges that connect vertices already in the same
    connected component. Hence the graph has no cycles — it is a union of trees.
    This graph-growing procedure, though motivated by the idea of merging clusters,
    is identically Kruskal’s algorithm. To produce a k-clustering of the objects, Kruskal’s
    algorithm is simply halted when there are k connected components, and the last k − 1
    edges are not added. This iterative merging procedure is also equivalent to computing the
    full minimal spanning tree, then deleting the k − 1 most expensive edges and taking the
    resulting k connected components to define a clustering C = {C1 , . . . Ck }.
    To see that this indeed produces k clusters with maximal spacing, note that the spacing
    of C = {C1 , . . . Ck } is the weight d∗ of the (k − 1)st most expensive edge (i.e., the next
    edge that would have been added) in the mimimal spanning tree. If C ′ = {C1′ , . . . Ck′ } is
    some other clustering, then there is some cluster Cr 6⊂ Cs′ such that there exist pi , pj ∈ Cr
    with pi ∈ Cs′ and pj ∈ Ct′ 6= Cs′ . Note that each edge on the path from pi to pj within Cr
    has weight ≤ d∗ . Let p′ be the first vertex along this path no longer in Cs′ and let p be the
    one just before p′ (i.e., still in Cs′ ). p and p′ are in different clusters of C ′ but d(p, p′ ) ≤ d∗ ,
    so the spacing of C ′ is no greater than that of C. The clustering C defined by the above
    procedure thus identifies a k-clustering with maximum spacing.

    INFO 295, 5 Oct 06

    You might also like