See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/2955136

CHAMELEON A hierarchical clustering algorithm using dynamic modeling

Article  in  Computer · September 1999

DOI: 10.1109/2.781637 · Source: IEEE Xplore

CITATIONS READS
1,648 1,395

3 authors:

George Karypis Eui-Hong Sam Han

University of Minnesota Twin Cities Persistent Systems Ltd
536 PUBLICATIONS   49,062 CITATIONS    33 PUBLICATIONS   4,042 CITATIONS   

SEE PROFILE SEE PROFILE

Vipin Kumar
University of Minnesota Twin Cities
636 PUBLICATIONS   49,838 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Understanding Climate Change: A Data Driven Approach View project

Sparse Tensor Factorization View project

All content following this page was uploaded by Vipin Kumar on 22 December 2013.

The user has requested enhancement of the downloaded file.

 To Appear in the IEEE Computer

C HAMELEON: A Hierarchical Clustering Algorithm

Using Dynamic Modeling

George Karypis Eui-Hong (Sam) Han Vipin Kumar

Department of Computer Science and Engineering
University of Minnesota
4-192 EECS Bldg., 200 Union St. SE
Minneapolis, MN 55455, USA
Technical Report #99–007
{karypis,han,kumar}@cs.umn.edu

Abstract
Clustering in data mining is a discovery process that groups a set of data such that the intracluster similarity
is maximized and the intercluster similarity is minimized. Existing clustering algorithms, such as K -means, PAM,
CLARANS, DBSCAN, CURE, and ROCK are designed to find clusters that fit some static models. These algorithms
can breakdown if the choice of parameters in the static model is incorrect with respect to the data set being clustered,
or if the model is not adequate to capture the characteristics of clusters. Furthermore, most of these algorithms
breakdown when the data consists of clusters that are of diverse shapes, densities, and sizes. In this paper, we present
a novel hierarchical clustering algorithm called C HAMELEON that measures the similarity of two clusters based on
a dynamic model. In the clustering process, two clusters are merged only if the inter-connectivity and closeness
(proximity) between two clusters are high relative to the internal inter-connectivity of the clusters and closeness of
items within the clusters. The merging process using the dynamic model presented in this paper facilitates discovery
of natural and homogeneous clusters. The methodology of dynamic modeling of clusters used in C HAMELEON is
applicable to all types of data as long as a similarity matrix can be constructed. We demonstrate the effectiveness
of C HAMELEON in a number of data sets that contain points in 2D space, and contain clusters of different shapes,
densities, sizes, noise, and artifacts. Experimental results on these data sets show that C HAMELEON can discover
natural clusters that many existing state-of-the art clustering algorithms fail to find.

Keywords: Clustering, data mining, dynamic modeling, graph partitioning, k-nearest neighbor graph.
1 Introduction
Clustering in data mining [SAD+ 93, CHY96] is a discovery process that groups a set of data such that the intracluster
similarity is maximized and the intercluster similarity is minimized [JD88, KR90, PAS96, CHY96]. These discovered
clusters can be used to explain the characteristics of the underlying data distribution, and thus serve as the foundation
for other data mining and analysis techniques. The applications of clustering include characterization of different
customer groups based upon purchasing patterns, categorization of documents on the World Wide Web [BGG+ 99a,
BGG+ 99b], grouping of genes and proteins that have similar functionality [HHS92, NRS+ 95, SCC+ 95, HKKM98],
grouping of spatial locations prone to earth quakes from seismological data [BR98, XEKS98], etc.
Existing clustering algorithms, such as K -means [JD88], PAM [KR90], CLARANS [NH94], DBSCAN [EKSX96],
CURE [GRS98], and ROCK [GRS99] are designed to find clusters that fit some static models. For example, K -means,
PAM, and CLARANS assume that clusters are hyper-ellipsoidal (or globular) and are of similar sizes. DBSCAN
assumes that all points within genuine clusters are density reachable 1 and points across different clusters are not.
Agglomerative hierarchical clustering algorithms, such as CURE and ROCK use a static model to determine the
most similar cluster to merge in the hierarchical clustering. CURE measures the similarity of two clusters based on
the similarity of the closest pair of the representative points belonging to different clusters, without considering the
internal closeness (i.e., density or homogeneity) of the two clusters involved. ROCK measures the similarity of two
clusters by comparing the aggregate inter-connectivity of two clusters against a user-specified static inter-connectivity
model, and thus ignores the potential variations in the inter-connectivity of different clusters within the same data set.
These algorithms can breakdown if the choice of parameters in the static model is incorrect with respect to the data set
being clustered, or if the model is not adequate to capture the characteristics of clusters. Furthermore, most of these
algorithms breakdown when the data consists of clusters that are of diverse shapes, densities, and sizes.
In this paper, we present a novel hierarchical clustering algorithm called C HAMELEON that measures the sim-
ilarity of two clusters based on a dynamic model. In the clustering process, two clusters are merged only if the
inter-connectivity and closeness (proximity) between two clusters are comparable to the internal inter-connectivity
of the clusters and closeness of items within the clusters. The merging process using the dynamic model presented
in this paper facilitates discovery of natural and homogeneous clusters. The methodology of dynamic modeling of
clusters used in C HAMELEON is applicable to all types of data as long as a similarity matrix can be constructed. We
demonstrate the effectiveness of C HAMELEON in a number of data sets that contain points in 2D space, and contain
clusters of different shapes, densities, sizes, noise, and artifacts.
The rest of the paper is organized as follows. Section 2 gives an overview of related clustering algorithms. Section 3
presents the limitations of the recently proposed state of the art clustering algorithms. We present our new clustering
algorithm in Section 4. Section 5 gives the experimental results. Section 6 contains conclusions and directions for
future work.

2 Related Work
In this section, we give a brief description of existing clustering algorithms.

1 A point p is density reachable from a point q, if they are connected by a chain of points such that each point has minimal number of data points,
including the next point in the chain, within a fixed radius [EKSX96].

1
2.1 Partitional Techniques
Partitional clustering attempts to break a data set into K clusters such that the partition optimizes a given crite-
rion [JD88, KR90, NH94, CS96]. Centroid-based approaches, as typified by K means [JD88] and ISODATA [BH64],
try to assign points to clusters such that the mean square distance of points to the centroid of the assigned cluster is
minimized. Centroid-based techniques are suitable only for data in metric spaces (e.g., Euclidean space) in which
it is possible to compute a centroid of a given set of points. Medoid-based methods, as typified by PAM (Partition-
ing Around Medoids) [KR90] and CLARANS [NH94], work with similarity data, i.e., data in an arbitrary similarity
space [GRG+ 99]. These techniques try to find representative points (medoids) so as to minimize the sum of the
distances of points from their closest medoid.
A major drawback of both of these schemes is that they fail for data in which points in a given cluster are closer
to the center of another cluster than to the center of their own cluster. This can happen in many natural clusters
[HKKM97, GRS99]; for example, if there is a large variation in cluster sizes (as in Figure 1 (a)) or when cluster
shapes are convex (as in Figure 1 (b)).

a) Clusters of widely differente sizes b) Clusters with convex shapes

Figure 1: Data sets on which centroid and medoid approaches fail.

2.2 Hierarchical Techniques

Hierarchical clustering algorithms produce a nested sequence of clusters, with a single all-inclusive cluster at the top
and single point clusters at the bottom. Agglomerative hierarchical algorithms [JD88] start with all the data points
as a separate cluster. Each step of the algorithm involves merging two clusters that are the most similar. After each
merge, the total number of clusters decreases by one. These steps can be repeated until the desired number of clusters
is obtained or the distance between two closest clusters is above a certain threshold distance.
There are many different variations of agglomerative hierarchical algorithms [JD88]. These algorithms primarily
differ in how they update the similarity between existing clusters and the merged clusters. In some methods [JD88],
each cluster is represented by a centroid or medoid of the points contained in the cluster, and the similarity between
two clusters is measured by the similarity between the centroids/medoids of the clusters. Like partitional techniques,
such as K -means and K -medoids, these method also fail on clusters of arbitrary shapes and different sizes.
In the single link method [JD88], each cluster is represented by all the data points in the cluster. The similarity
between two clusters is measured by the similarity of the closest pair of data points belonging to different clusters. Un-
like the centroid/medoid based methods, this method can find clusters of arbitrary shape and different sizes. However,
this method is highly susceptible to noise, outliers, and artifacts.
CURE [GRS98] has been proposed to remedy the drawbacks of both of these methods while combining their
advantages. In CURE, instead of using a single centroid to represent a cluster, a constant number of representative

2
points are chosen to represent a cluster. The similarity between two clusters is measured by the similarity of the closest
pair of the representative points belonging to different clusters. New representative points for the merged clusters are
determined by selecting a constant number of well scattered points from all the data points and shrinking them towards
the centroid of the cluster according to a shrinking factor. Unlike centroid/medoid based methods, CURE is capable of
finding clusters of arbitrary shapes and sizes, as it represents each cluster via multiple representative points. Shrinking
the representative points towards the centroid helps CURE in avoiding the problem of noise and outliers present in the
single link method. The desirable value of the shrinking factor in CURE is dependent upon cluster shapes and sizes,
and amount of noise in the data.
In some agglomerative hierarchical algorithms, the similarity between two clusters is captured by the aggregate
of the similarities (i.e., interconnectivity) among pairs of items belonging to different clusters. The rationale for this
approach is that subclusters belonging to the same cluster will tend to have high interconnectivity. But the aggregate
inter-connectivity between two clusters depends on the size of the clusters involved, and in general pairs of larger
clusters will have higher inter-connectivity. Hence, many such schemes normalize the aggregate similarity between
a pair of clusters with respect to the expected inter-connectivity of the clusters involved. For example, the widely
used group-average method [JD88] assumes fully connected clusters, and thus scales the aggregate similarity between
two clusters by n × m, where n and m are the size of the two clusters, respectively. ROCK [GRS99], a recently
developed agglomerative algorithm that operates on a derived similarity graph, scales the aggregate inter-connectivity
with respect to a user-specified inter-connectivity model.
Most of the algorithms discussed above work implicitly or explicitly with the n × n similarity matrix such that (i, j )
element of the matrix represents the similarity between i t h and j t h data items. Some algorithms derive a new similarity
matrix using the original matrix [JP73, GK78, JD88, GRS99], and then apply one of the existing techniques on this
derived similarity matrix. In many cases, the new derived similarity matrix is just a sparsified version of this original
similarity matrix from which certain entries (e.g., those whose value is below a threshold) have been deleted. In other
cases, the derived similarity matrix has entirely different values [JP73, GK78, GRS99]. The sparsified derived matrix
can help eliminate/reduce noise from the data, and substantially reduce the execution time of many algorithms. In some
cases, it can also provide a better model of similarities for the problem domain. For example, mutual shared method
presented in [JP73] helps remove noise and outliers and is shown to provide a better model to capture similarities
among transactions in [GRS99].
A sparse similarity matrix can be represented by a sparse graph, and tightly connected clusters of this graph can be
found by divisive hierarchical clustering algorithms such as those based upon minimal spanning tree (MST) [JD88] or
graph-partitioning algorithms [KK98b, KK99a]. MST-based algorithms are highly susceptible to noise and artifacts
just like the single link method. Graph-partitioning based methods are much more robust, but they tend to break
genuine clusters if there is a large variations in cluster sizes.

3 Limitations of Existing Hierarchical Schemes

A major limitation of existing agglomerative hierarchical schemes such as the Group Averaging Method [JD88],
ROCK [GRS99], and CURE [GRS98] is that the merging decisions are based upon static modeling of the clusters to
be merged. In other words, these schemes fail to take into account special characteristics of individual clusters, and
thus can make incorrect merging decisions when the underlying data does not follow the assumed model, or when noise
is present. For example, consider the four sub-clusters of points in 2D shown in Figure 2. The selection mechanism of
CURE (and of the single link method) will prefer merging clusters (a) and (b) over merging clusters (c) and (d), since

3
the minimum distances between the representative points of (a) and (b) will be smaller than those for clusters (c) and
(d). But clusters (c) and (d) are better candidates for merging because the minimum distances between the boundary
points of (c) and (d) are of the same order as the average of the minimum distances of any points within these clusters
to other points. Hence, merging (c) and (d) will lead to a more homogeneous and natural cluster than merging (a) and
(b).

(a) (b) (c) (d)

Figure 2: Example of clusters for merging choices.

In agglomerative schemes based upon group averaging [JD88] and related schemes such as ROCK, connectivity
among pairs of clusters is scaled with respect to the expected connectivity between these clusters. However, the key
limitation of all such schemes is that they assume a static, user supplied inter-connectivity model, which is inflexible
and can easily lead to wrong merging decisions when the model under- or over-estimates the inter-connectivity of the
data set or when different clusters exhibit different inter-connectivity characteristics. Although some schemes allow
the connectivity to be different for different problem domains (e.g., ROCK [GRS99]), it is still the same for all clusters
irrespective of their densities and shapes. Consider the two pairs of clusters shown in Figure 3, where each cluster
is depicted by a sparse graph where nodes indicate data items and edges represent that their two vertices are similar.
The number of items in all four clusters is the same. Let us assume that in this example all edges have equal weight
(i.e., they represent equal similarity). Then both ROCK selection mechanism (irrespective of the assumed model of
connectivity) and the group averaging method will select pair {(c),(d)} for merging, whereas the pair {(a),(b)} is a
better choice.

(a) (b)

(c) (d)

Figure 3: Example of clusters for merging choices.

The selection mechanism in CURE (and related algorithms such as single link method [JD88]) considers only the
minimum distance between the representative points of two clusters, and does not consider the aggregate intercon-
nectivity among the two clusters. Similarly, the selection mechanism of algorithms such as ROCK only considers
the aggregate inter-connectivity across the pairs of clusters (appropriately scaled by the expected value of the inter-
connectivity), but ignores the value of the strongest edge (or edges) across clusters. However, by looking at only one

4
of these two characteristics, these algorithm can easily select to merge the wrong pair of clusters. For instance, as the
example in Figure 4 illustrates, an algorithm that focuses only on the closeness of two clusters will incorrectly prefer
to merge clusters (c) and (d) over clusters (a) and (b). Similarly, as the example in Figure 5 illustrates, an algorithm
that focuses only on the inter-connectivity of two clusters will incorrectly prefer to merge cluster (a) with cluster (c)
rather than with (b). (Here we assume that the aggregate interconnectivity between items in clusters (a) and (c) is
greater than that between items in clusters (a) and (b). However, the border points of cluster (a) are much closer than
those of (b) than to those of (c).)

(a) (b) (c) (d)

Figure 4: Example of clusters for merging choices.

(a) (b)

(c)

Figure 5: Example of clusters for merging choices.

In summary, there are two major limitations of the agglomerative mechanisms used in existing schemes. First,
these schemes do not make use of information about the nature of individual clusters being merged. Second, one
set of schemes (CURE and related schemes) ignore the information about the aggregate interconnectivity of items
in two clusters, whereas the other set of schemes (ROCK, the group averaging method, and related schemes) ignore
information about the closeness of two clusters as defined by the similarity of the closest items across two clusters.
In the following section, we present a novel scheme that addresses both of these limitations.

4 C HAMELEON: Clustering Using Dynamic Modeling

4.1 Overview
In this section we present C HAMELEON, a new clustering algorithm that overcomes the limitations of existing ag-
glomerative hierarchical clustering algorithms discussed in Section 3. Figure 6 provides an overview of the overall
approach used by C HAMELEON to find the clusters in a data set.
C HAMELEON operates on a sparse graph in which nodes represent data items, and weighted edges represent sim-
ilarities among the data items. This sparse graph representation of the data set allows C HAMELEON to scale to large
data sets and to operate successfully on data sets that are available only in similarity space [GRG+ 99] and not in
metric spaces [GRG+ 99]. C HAMELEON finds the clusters in the data set by using a two phase algorithm. During the
first phase, C HAMELEON uses a graph partitioning algorithm to cluster the data items into a large number of relatively
small sub-clusters. During the second phase, it uses an agglomerative hierarchical clustering algorithm to find the
genuine clusters by repeatedly combining together these sub-clusters.

5
 K-nearest Neighbor Graph Final Clusters
Data Set

Construct a
Sparse Graph Partition the Graph Merge Partitions

Figure 6: Overall framework C HAMELEON.

The key feature of C HAMELEON’s agglomerative hierarchical clustering algorithm is that it determines the pair of
most similar sub-clusters by taking into account both the inter-connectivity as well as the closeness of the clusters;
and thus it overcomes the limitations discussed in Section 3 that result from using only one of them. Furthermore,
C HAMELEON uses a novel approach to model the degree of inter-connectivity and closeness between each pair of
clusters that takes into account the internal characteristics of the clusters themselves. Thus, it does not depend on a
static user supplied model, and can automatically adapt to the internal characteristics of the clusters being merged.
In the rest of this section we provide details on how to model the data set, how to dynamically model the similarity
between the clusters by computing their relative inter-connectivity and relative closeness, how graph partitioning is
used to obtain the initial fine-grain clustering solution, and how the relative inter-connectivity and relative closeness
are used to repeatedly combine together the sub-clusters in a hierarchical fashion.

4.2 Modeling the Data

Given a similarity matrix, many methods can be used to find a graph representation [JP73, GK78, JD88, GRS99]. In
fact, modeling data items as a graph is very common in many hierarchical clustering algorithms. For example, ag-
glomerative hierarchical clustering algorithms based on single link, complete link, or group averaging method [JD88]
operate on a complete graph. ROCK [GRS99] first constructs a sparse graph from a given data similarity matrix using
a similarity threshold and the concept of shared neighbors, and then performs a hierarchical clustering algorithm on
the sparse graph. CURE [GRS98] also implicitly employs the concept of a graph. In CURE, when cluster representa-
tive points are determined, a graph containing only these representative points is implicitly constructed. In this graph,
edges only connect representative points from different clusters. Then the closest edge in this graph is identified and
the clusters connected by this edge is merged.
C HAMELEON’s sparse graph representation of the data items is based on the commonly used k-nearest neighbor
graph approach. Each vertex of the k-nearest neighbor graph represents a data item, and there exists an edge between
two vertices, if data items corresponding to either of the nodes is among the k-most similar data points of the data
point corresponding to the other node. Figure 7 illustrates the 1-, 2-, and 3-nearest neighbor graphs of a simple data
set. Note that since C HAMELEON operates on a sparse graph, each cluster is nothing more than a sub-graph of the
original sparse graph representation of the data set.
There are several advantages of representing data using a k-nearest neighbor graph G k . Firstly, data points that are
far apart are completely disconnected in the G k . Secondly, G k captures the concept of neighborhood dynamically.
The neighborhood radius of a data point is determined by the density of the region in which this data point resides.
In a dense region, the neighborhood is defined narrowly and in a sparse region, the neighborhood is defined more
widely. Compared to the model defined by DBSCAN [EKSX96] in which a global neighborhood density is specified,
G k captures more natural neighborhood. Thirdly, the density of the region is recorded as the weights of the edges. The
edge weights of dense regions in G k (with edge weights representing similarities) tend to be large and the edge weights

6
(a) Original Data in 2D (b) 1-nearest neighbor graph (c) 2-nearest neighbor graph (d) 3-nearest neighbor graph

Figure 7: k-nearest graphs from an original data in 2D.

of sparse regions tend to be small. As the consequence, a min-cut bisection of the graph represents the interface layer
of sparse region of the graph. Finally, G k provides a computational advantage over a full graph in many algorithms
operating on graphs, including graph partitioning and partitioning refinement algorithms.

4.3 Modeling the Cluster Similarity

To address the limitations of agglomerative schemes discussed in Section 3, C HAMELEON determines the similarity
between each pair of clusters Ci and C j by looking both at their relative inter-connectivity R I (Ci , C j ) and their
relative closeness RC(Ci , C j ). C HAMELEON’s hierarchical clustering algorithm selects to merge the pair of clusters
for which both R I (Ci , C j ) and RC(Ci , C j ) are high; i.e., it selects to merge clusters that are well inter-connected as
well as close together with respect to the internal inter-connectivity and closeness of the clusters. By selecting clusters
based on both of these criteria, C HAMELEON overcomes the limitations of existing algorithms that look either at the
absolute inter-connectivity or absolute closeness. For instance, in the examples shown in Figures 4 and 5 and discussed
in Section 3, C HAMELEON will select to merge the correct pair of clusters.
In the remaining of this section we describe how the relative inter-connectivity and relative closeness is computed
for a pair of clusters.

Relative Inter-Connectivity The relative inter-connectivity between a pair of clusters Ci and C j is defined as
the absolute inter-connectivity between Ci and C j normalized with respect to the internal inter-connectivity of the two
clusters Ci and C j . The absolute inter-connectivity between a pair of clusters Ci and C j is defined to be as the sum
of the weight of the edges that connect vertices in Ci to vertices in C j . This is essentially the edge-cut of the cluster
containing both Ci and C j such that the cluster is broken into Ci and C j . We denote this by EC{Ci ,C j } . The internal
inter-connectivity of a cluster Ci can be easily captured by the size of its min-cut bisector ECCi (i.e., the weighted sum
of edges that partition the graph into two roughly equal parts). Recent advances in the graph-partitioning technology
has made it possible to find such bisector quite efficiently [KK98b, KK99a].
Thus the relative inter-connectivity between a pair of clusters Ci and C j is given by

|EC{Ci ,C j } |
R I (Ci , C j ) = |ECCi |+|ECC j |
, (1)
2

which normalizes the absolute inter-connectivity with the average internal inter-connectivity of the two clusters.
By focusing on the relative inter-connectivity between clusters, C HAMELEON can overcome the limitations of
existing algorithms that use static inter-connectivity models. For instance, in the example shown in Figure 3 that
was discussed in Section 3, C HAMELEON will correctly prefer to merge clusters (a) and (b) over clusters (c) and

7
(d), because the relative inter-connectivity between clusters (a) and (b) is higher than the relative inter-connectivity
between clusters (c) and (d), even though the later pair of clusters have a higher absolute inter-connectivity. Thus, the
relative inter-connectivity is able to take into account differences in shapes of the clusters (as in Figure 3) as well as
differences in degree of connectivity of different clusters.

Relative Closeness The relative closeness between a pair of clusters Ci and C j is defined as the absolute close-
ness between Ci and C j normalized with respect to the internal closeness of the two clusters Ci and C j . The absolute
closeness between a pair of clusters can be captured in a number of different ways. Many existing schemes, capture
this closeness by focusing on the pair of points between all the points (or representative points [GRS98]) from Ci and
C j that are closest. A key drawback of these schemes is that by relying only on a single pair of points, they are less
tolerant to outliers and noise. For this reason, C HAMELEON measures the closeness of two clusters by computing the
average similarity between the points in Ci that are connected to points in C j . Since these connections are determined
using the k-nearest neighbor graph, their average strength provides a very good measure of the affinity between the
data items along the interface layer of the two sub-clusters, and at the same time is tolerant to outliers and noise.
Note that this average similarity between the points from the two clusters is equal to the average weight of the edges
connecting vertices in Ci to vertices in C j .
The internal closeness of each cluster Ci can also be measured in a number of different ways. One possible approach
is to look at all the edges connecting vertices in Ci (i.e., edges that are internal to the cluster), and compute the internal
closeness of a cluster as the average weight of these edges. One can argue that in a hierarchical clustering setting,
the edges used for agglomeration early on are stronger than those used in later stages. Hence, average weights of the
edges on the internal bisection of Ci and C j will tend to be smaller than the average weight of all the edges in these
clusters. But the average weight of these edges is a better indicator of the internal closeness of these clusters.
Hence in C HAMELEON, the relative closeness between a pair of clusters Ci and C j is computed as,

S EC{Ci ,C j }
RC(Ci , C j ) = |C j |
, (2)
|Ci |
|Ci |+|C j | S ECCi + |Ci |+|C j | S ECC j

where S ECCi and S ECC j are the average weights of the edges that belong in the min-cut bisector of clusters Ci and C j ,
respectively, and S EC{Ci ,C j } is the average weight of the edges that connect vertices in Ci to vertices in C j . Also note
that a weighted average of the internal closeness of clusters Ci and C j is used to normalize the absolute closeness of
the two clusters, that favors the absolute closeness of cluster that contains the larger number of vertices.
By focusing on the relative closeness between clusters, C HAMELEON can overcome the limitations of existing
algorithms that look only at the absolute closeness. For instance, in the example shown in Figure 2 that was discussed
in Section 3, C HAMELEON will correctly prefer to merge the clusters (c) and (d) over the clusters (a) and (b). This
is because, the relative closeness of clusters (c) and (d) is higher than the relative closeness between clusters (a)
and (b), even though the later pair of clusters have a higher absolute closeness. Thus, by looking at the relative
closeness, C HAMELEON correctly prefers to merge clusters whose resulting cluster exhibits a uniformity in the degree
of closeness between the items in the cluster. Also note that the relative closeness between two clusters is in general
smaller than one, because the edges that connect vertices in different clusters have a smaller weight.

8
4.4 C HAMELEON : A Two-phase Clustering Algorithm
The dynamic framework for modeling the similarity between clusters discussed in Section 4.3 can be only applied
when each cluster contains a sufficiently large number of vertices (i.e., data items). This is because in order to compute
the relative inter-connectivity and relative closeness of clusters, C HAMELEON needs to compute the internal inter-
connectivity and closeness of each cluster. Both of which cannot be accurately calculated for clusters containing only
a few data points. For this reason, C HAMELEON uses an algorithm that consists of two distinct phases. The purpose
of the first phase is to cluster the data items into a large number of sub-clusters that contain a sufficient number of
items to allow dynamic modeling. The purpose of the second phase, is to discover the genuine clusters in the data
set by using the dynamic modeling framework to merge together these sub-clusters in a hierarchical fashion. In the
remainder of this section, we present the algorithms used for these two phases of CHAMELEON.

Phase I: Finding Initial Sub-clusters C HAMELEON finds the initial sub-clusters using a graph partitioning
algorithm to partition the k-nearest neighbor graph of the data set into a large number of partitions such that the edge-
cut, i.e., the sum of the weight of the edges that straddle partitions, is minimized. Since each edge in the k-nearest
neighbor graph represents the similarity among data points, a partitioning that minimizes the edge-cut effectively
minimizes the relationship (affinity) among data points across the resulting partitions. The underlying assumption is
that links within clusters will be stronger and more plentiful than links across clusters. Hence, the data in each partition
are highly related to other data items in the same partition.
Recent research on graph partitioning has lead to the development of fast and accurate algorithms that are based
on the multilevel paradigm [KK99a, KK99b]. Extensive experiments on graphs arising in many application domains
have shown that multilevel graph partitioning algorithms are very effective in capturing the global structure of the
graph and are capable of computing partitionings that have a very small edge-cut. Hence, when used to partition the
k-nearest neighbor graph, they are very effective in finding the natural separation boundaries of clusters. For example,
Figure 8 shows the two clusters produced by applying a multilevel graph partitioning algorithm on the k-nearest-
neighbor graphs for two spatial data sets. As we can see from this figure, the partitioning algorithm is very effective in
finding the low-density separating region in the first example, and the small connecting region in the second example.

(a) (b)

Figure 8: An example of the bisections produced by multilevel graph partitioning algorithms on two spatial data sets. (a) The
partitioning algorithm cuts through the sparse region. (b) The partitioning algorithms cuts through a small connecting region.

C HAMELEON utilizes such multilevel graph partitioning algorithms to find the initial sub-clusters. In particular,
it uses the graph partitioning algorithm that is part of the hMETIS library [KK98a]. hMETIS has been shown [KK98c,
KK99b, Alp98] to quickly produce high-quality partitionings for a wide range of unstructured graphs and hypergraphs.

9
In C HAMELEON we primarily use hMETIS to split a cluster Ci into two sub-clusters CiA and CiB such that the edge-cut
between CiA and CiB is minimized and each one of these sub-clusters contains at least 25% of the nodes in Ci . Note
that this last requirement, often referred to as the balance constraint, is an integral part of using a graph partitioning
approach to find the sub-clusters. hMETIS is effective in operating within the allowed balance constraints to find a
bisection that minimizes the edge-cut. However, this balance constraint can force hMETIS to break a natural cluster.
C HAMELEON obtains the initial set of sub-clusters as follows. It initially starts with all the points belonging to the
same cluster. It then repeatedly selects the largest sub-cluster among the current set of sub-clusters and uses hMETIS
to bisect. This process terminates when the larger sub-cluster contains fewer than a specified number of vertices, that
we will refer to it as M IN S IZE. The M IN S IZE parameter essentially controls the granularity of the initial clustering
solution. In general, M IN S IZE should be set to a value that is smaller than the size of most of the clusters that we
expect to find in the data set. At the same time, M IN S IZE should be sufficiently large such that most of the sub-clusters
contain a sufficiently large number of nodes to allow us to evaluate the inter-connectivity and closeness of the items in
each sub-cluster in a meaningful fashion. For most of the data sets that we encountered, setting M IN S IZE to about 1%
to 5% of the overall number of data points worked fairly well.

Phase II: Merging Sub-Clusters using a Dynamic Framework As soon as the fine-grain clustering solution
produced by the partitioning-based algorithm of the first phase is found, C HAMELEON then switches to an agglomer-
ative hierarchical clustering that combines together these small sub-clusters. As discussed in Section 2, the key step
of agglomerative hierarchical algorithm is that of finding the pair of sub-clusters that are the most similar.
C HAMELEON’s agglomerative hierarchical clustering algorithm utilizes the dynamic modeling framework dis-
cussed in Section 4.3 to select the most similar pairs of clusters by looking both at their relative inter-connectivity
and their relative closeness. There are many ways to develop an agglomerative hierarchical clustering algorithm that
takes into account both of these measures. Two different schemes have been implemented in C HAMELEON.
The first scheme merges only those pairs of clusters whose relative inter-connectivity and relative closeness are both
above some user specified threshold TR I and TRC , respectively. In this approach, C HAMELEON visits each cluster Ci ,
and checks to see if any one of its adjacent clusters C j satisfy the following two conditions:

R I (Ci , C j ) ≥ TR I and RC(Ci , C j ) ≥ TRC . (3)

If more than one of the adjacent clusters satisfy the above conditions, then CHAMELEON selects to merge Ci with the
cluster that it is most connected to; i.e., it selects the cluster C j such that the absolute inter-connectivity between these
two clusters is the highest. Once every cluster has been given the opportunity to merge with one of its adjacent clusters,
the combinations that have been selected are performed, and the entire process is repeated. Note that this algorithm is
different than traditional hierarchical clustering algorithms, as it allows multiple pairs of clusters to be merged together
at any given iteration. The parameters TR I and TRC can be used to control the characteristics of the desired clusters.
In particular, the parameter TR I allows us to control the variability in the degree of inter-connectivity of the items
in the cluster. The parameter TRC allows us to control the uniformity of the similarity among items that belong to a
particular cluster. Depending on the choice of the TR I and TRC parameters, C HAMELEON’s merging algorithm may
reach a point from which it cannot proceed any further because none of the adjacent clusters satisfy the two conditions
of Equation 3. At this point we have the choice of either terminating the algorithm and output the current clustering
as the solution or try to merge additional pairs of clusters by successively relaxing the two parameters, possibly at
different rates.

10
 The second scheme implemented in C HAMELEON uses a function to combine the relative inter-connectivity and
relative closeness, and then selects to merge the pair of clusters that maximizes this function. Since our goal is to
merge together pairs for which both the relative inter-connectivity and the relative closeness are high, a natural way of
defining such a function is to take their product. That is, select the pair of clusters Ci and C j to merge that maximize
R I (Ci , C j ) ∗ RC(Ci , C j ). This formula gives an equal importance to both of these parameters. However, quite often
we may prefer clusters that give a higher preference to one of these two measures. For this reason, CHAMELEON
selects the pair of clusters that maximizes

R I (Ci , C j ) ∗ RC(Ci , C j )α , (4)

where α is a user specified parameter. If α > 1, then C HAMELEON gives a higher importance to the relative closeness,
and when α < 1, it gives a higher importance on the relative inter-connectivity. In the experimental results presented
in Section 5 we used this second approach as it allows us to easily generate the entire dendrogram for the hierarchical
clustering.

4.5 Performance Analysis

The overall computational complexity of C HAMELEON depends on the amount of time it requires to construct the
k-nearest neighbor graph and the amount of time it requires to perform the two phases of the clustering algorithm.
The amount of time required to compute the k-nearest neighbor graph depends on the dimensionality of the under-
lying data set. In particular, for low-dimensional data sets, algorithms based on k − d trees [Sam90] can be used to
quickly compute the k nearest neighbors. It has been shown that for n items, the average cost of inserting, as well
as the expected k-nearest neighbor search time is O(log n) [FBF77], leading to an overall complexity of O(n log n).
However, for high dimensional data sets, schemes based on k − d trees are not applicable [BBKK97, BBK98]. For
such data sets, the amount of time required to find the k-nearest neighbors of a data item is O(n), leading to an overall
complexity of O(n 2 ).
The amount of time required by C HAMELEON’s two-phase clustering algorithm depends on the number m of
initial sub-clusters produced by the graph partitioning algorithm used in the first phase. To simplify the analysis,
we will assume that (i) each initial sub-cluster has the same number of nodes n/m, and (ii) during each successive
merging step, C HAMELEON selects to merge only a single pair of clusters. Moreover, our analysis will be focused
on C HAMELEON’s second scheme for combining the relative inter-connectivity and relative closeness described in
Section 4.4. However, the overall complexity is similar for the first scheme as well.
The amount of time required by the the Phase I of C HAMELEON depends on the amount of time required by
hMETIS. Given a graph G = (V, E), hMETIS requires O(|V | + |E|) [KK98c, KK99b] time to compute a bisection.
Since C HAMELEON operates on the k-nearest neighbor graph, |E| = O(|V |); thus, the computational complexity
of hMETIS is O(|V |). C HAMELEON’s Phase I algorithm obtains m clusters by repeatedly partitioning successively
smaller graphs; hence, its overall computational complexity is O(n log(n/m)) which is bounded by O(n log n). Note
that one can potentially use a faster partitioning algorithm to obtain the initial m clusters in time O(n + m log m) using
the multilevel m-way partitioning algorithm described in [KK99b].
The amount of time required by the second phase depends on (a) the amount of time needed to compute the internal
inter-connectivity and internal closeness for each initial as well as intermediate cluster, and (b) the amount of time
needed to select the most similar pair of clusters to merge. Since the internal inter-connectivity and internal closeness
for a particular cluster is computed by bisecting the corresponding k-nearest neighbor sub-graph of the cluster, its

11
complexity is proportional to the number of items in each cluster. In particular, the amount of time required to bisect
each one of the initial m clusters is O(n/m), leading to an overall complexity of O(n). Next, during each of the
merging steps, C HAMELEON needs to bisect the resulting cluster, and there are a total of m − 1 such steps (i.e.,
until all the initial sub-clusters have been merged together). The worst case complexity is obtained when the merging
algorithm repeatedly selects the same cluster and merges it with another; i.e., it grows a single large cluster. This
corresponds to the worst case because during each merging step, the algorithm needs to bisect a cluster that has the
largest possible number of data items. In this case, the amount of time required to bisect the m − 1 intermediate
P
sub-clusters is m−1
i=2 (i × n/m) which is O(nm).
The overall amount of time required to find the most similar pair of clusters is O(m 2 log m) by using a heap-based
priority queue. In the worst case, the initial clustering solution can be such that each cluster is connected to all the
remaining clusters. In this case, it takes O(m 2 log m) time to insert the similarity of the O(m 2 ) possible pairs of
sub-clusters into the priority queue. Now, during each merging step, the pair residing at the top of the priority queue
is selected, and the similarity of recently combined cluster to the remaining sub-clusters is updated. Each of these
update operations requires O(m log m) time, leading to an overall complexity of O(m 2 log m), as a total of m − 1 such
updates needs to be performed (one for each pair of clusters that gets merged).
Thus, the overall complexity of C HAMELEON’s two-phase clustering algorithm is O(nm + n log n + m 2 log m).

5 Experimental Results
In this section, we present experimental evaluation of C HAMELEON, and compare its performance with a publicly
available version of DBSCAN and a locally implemented version of CURE. Even though CHAMELEON is applicable
to any data set for which a similarity matrix is available (or can be constructed), we chose to perform evaluation
for data sets containing points in two dimensional space for two reasons. First, similar data sets have been used to
evaluate the performance of other state-of-the art algorithms such as DBSCAN and CURE. Second, clusters in 2D
data sets are easy to visualize, making the comparison of different schemes much easier. We do not report results of
ROCK [GRS99] (and other interconnectivity based agglomerative schemes such as group averaging method [JD88]),
as they tend to perform worse than algorithms such as CURE on metric space data sets. Many of these results are
available at URL http://www.cs.umn.edu/∼han/chameleon.html.

5.1 Data Sets

We experimented with five different data sets containing points in two dimensions whose geometric shape are shown
in Figure 9. The first data set, DS1, has five clusters that are of different size, shape, and density, and contains noise
points as well as special artifacts. The second data set, DS2, contains two clusters that are close to each other and
different regions of the clusters have different densities. The third data set, DS3, has six clusters of different size,
shape, and orientation, as well as random noise points and special artifacts such as streaks running across clusters. The
fourth data set, DS4, has eight clusters of different shape, size, and orientation, some of which are inside the space
enclosed by other clusters. Moreover, DS4 also contains random noise and special artifacts, such as a collection of
points forming vertical streaks. Finally, the fifth data set, DS5, has eight clusters of different shape, size, density, and
orientation, as well as random noise. A particularly challenging feature of this data set is that clusters are very close to
each other and they have different densities. The size of these data sets ranges from 6,000 to 10,000 points, and their
exact size is indicated in Figure 9. Note that DS1 was obtained from [GRS98], whereas we synthetically generated the
remaining data sets.

12
 DS1: 8000 points DS2: 6000 points DS3: 8000 points

DS4: 10000 points DS5: 8000 points

Figure 9: The five data sets used in our experiments.

5.2 Qualitative Comparison

C HAMELEON To cluster a data set using C HAMELEON, we need to specify the following parameters: the value of k
for computing the k-nearest neighbor graph, the value of M IN S IZE for the Phase I of the algorithm, and the choice of
scheme for combining relative inter-connectivity and relative closeness and associated parameters. In the experiments
presented in this section, we used the same set of parameter values for all five data sets. In particular, we used k = 10,
M IN S IZE = 2.5% of the total items in the data set, and used the second scheme for combining RI and RC, and used
α = 2.0 in Equation 4 for combining relative inter-connectivity and relative closeness of each pair of clusters. We
also performed a parameter study to determine the sensitivity of C HAMELEON on the above set of parameters by using
k = {5, 10, 15, 20}, M IN S IZE = {2%, 3%, 4%} and α = {1.5, 2.0, 2.5, 3.0}. Our results (not shown here) show that
C HAMELEON is not very sensitive on the above choice of parameters, and it was able to discover the correct clusters
for all of these combinations of values for k, M IN S IZE, and α.
Figure 10 shows the clusters found by C HAMELEON for each one of the five data sets. The points in the different
clusters are represented using a combination of different colors and different glyphs. As a result, points that belong
to the same cluster have both the same color as well as their points are drawn using the same glyph. For example,
in the clustering solution shown for DS4, there are two clusters that have cyan color (one contains the points in the
region between the two circles inside the ellipse, and the other contains the points that form a line between the two
horizontal bars and the ’c’ shaped cluster), and there are two clusters that have a dark blue color (one corresponds
to the upside-down ’c’ shaped cluster and the other corresponds to the circle inside the candy-cane); however, their
points are represented using different glyphs (bells and squares for the first pair, and squares and bells for the second
pair), so they denote different clusters.
Since C HAMELEON is hierarchical in nature, it creates a dendrogram of possible clustering solutions at different
levels of granularity. The clustering solutions shown in Figure 10 correspond to the earliest point in the agglomerative
process in which C HAMELEON was able to find the genuine clusters in the data set. That is, they correspond to the
lowest level of the dendrogram at which the genuine clusters in the data set have been identified and each one has been
placed together in one cluster. As a result, the number of clusters shown in Figure 10 for each one of the data sets can

13
be larger than the number of genuine clusters, and these additional clusters contain points that are outliers.
Looking at Figure 10, we can see that C HAMELEON is able to correctly identify the genuine clusters in all five
data sets. In the case of DS1, C HAMELEON finds six clusters, five of which correspond to the genuine clusters in the
data set, and the sixth one (shown with brown-colored ’*’ glyphs) corresponds to outlier points connecting the two
ellipsoid clusters. In the case of DS2, C HAMELEON finds two clusters, each one corresponding to a genuine cluster in
the data set. In the case of DS3, C HAMELEON finds eleven clusters, out of which six of them correspond to the genuine
clusters in the data set, and the rest contain outliers. In the case of DS4, C HAMELEON also finds eleven clusters, out of
which nine of them correspond to the genuine clusters, and the rest contain outlier points. Finally, in the case of DS5,
C HAMELEON finds eight clusters, each one corresponding to a genuine cluster in the data set. As these experiment
illustrate C HAMELEON is very effective in finding clusters of arbitrary shape, density, and orientation, and is tolerant
to outlier points, as well as artifacts such as streaks running across clusters.

CURE We evaluated the performance of C HAMELEON against CURE (described in Section 2) which has been
shown to be effective in finding clusters in two dimensional point data sets [GRS98]. CURE was able to find the right
clusters for DS1 and DS2, but it failed to find the right clusters on the remaining three data sets. Figure 11 shows the
results obtained by CURE for each one of the DS3, DS4, and DS5 data sets. Since CURE is also hierarchical clustering
algorithm, it also produces a dendrogram of possible clustering solutions at different levels of granularity. For each one
of the data sets, Figure 11 shows two different clustering solutions containing different number of clusters. The first
clustering solution (first column of Figure 11) corresponds to the earliest point in the agglomerative process in which
CURE merges together sub-clusters that belong to two different genuine clusters. As we can see from Figure 11, in
the case of DS3, CURE selects the wrong pair of clusters to merge it together when going from 18 down to 17 clusters,
resulting in the red-colored sub-cluster which contains portions of the two π-shaped clusters. Similarly, in the case
of DS4, CURE makes a mistake when going from 26 down to 25 clusters, as it selects to merge together one of the
circles inside the ellipse with a portion of the ellipse. Finally, in the case of DS5, CURE also makes a mistake when
going from 26 down to 25 clusters by merging together the small circular cluster with a portion of the upside-down
’Y’-shaped cluster. The second clustering solution corresponds to solutions that contain as many clusters as those
discovered by C HAMELEON. These solutions are considerably worse than the first set of solutions (especially for DS4
and DS5), indicating that the merging scheme used by CURE performs multiple mistakes.
For the results shown in Figure 11 experiments, the shrinking factor is 0.3 and the number of representative points
is 10, which are the default values recommended in [GRS98]. We also performed experiments with shrinking factor
varying from 0.1 to 0.9 and the number of representative points varying from 10 to 100. These experiments showed
similar trends as shown in Figure 11. Furthermore, to facilitate fair comparisons, we also removed the noise as
suggested in [GRS98], by identifying “slowly” growing clusters as noise point and removed them. For comparison
purposes we reassigned these noisy data points back to the final clusters using the assignment method discussed in
[GRS98], i.e., noise points are assigned to the cluster with the closest representative points. Note that these assignments
did not affect the overall clustering results.

DBSCAN DBSCAN [EKSX96] is a well-known spatial clustering algorithm that has been shown to find clusters
of arbitrary shapes. DBSCAN defines a cluster to be a maximum set of density-connected points. Every core point
in a cluster must have at least a minimum number of points (MinPts) within a given radius (Eps). DBSCAN can find
arbitrary shape of clusters if the right density of the clusters can be determined in a priori and the density of clusters is
uniform.

14
 DBSCAN finds the right clusters on data sets DS3 and DS4 as long as it is supplied the right combination of Eps
and MinPts. If MinPts is fixed to 4 (default value specified in [EKSX96]), then the algorithm works fine as long as
Eps is within the range (5.0,5.4) for DS3 and (5.7,6.1) for DS4. However, it fails to perform well on DS1, DS2, and
DS5, as these data sets contain clusters of different densities.
Figure 12 shows the clusters found by DBSCAN for DS1 and DS2 for different values of the Eps parameter.
Following the recommendation of [EKSX96], the MinPts was fixed to 4 and Eps was changed in these experiments.
The clusters produced for DS1 illustrate that DBSCAN cannot effectively find clusters of different density. In the first
clustering solution (Figure 12(a)), when Eps = 0.5, DBSCAN puts the two ellipses into the same cluster, because the
outlier points connecting satisfy the density requirements as dictated by the Eps and MinPts parameters. These clusters
can be separated by decreasing the value of Eps as was done in the clustering solution shown in Figure 12(b), for which
Eps=0.4. However, DBSCAN keeps the ellipses together, but now it has fragmented the lower density cluster into a
large number of small sub-clusters. Our experiments has shown that DBSCAN exhibits similar characteristics on DS5.
The clusters produced for DS2 illustrate that DBSCAN cannot effectively find clusters that their internal density varies.
The sequence of the three clustering solutions (Figure 12(c)–(e)) for decreasing values of the Eps parameter illustrates
that as we decrease Eps in hope of separating the two clusters, the natural clusters in the data set are fragmented
into a large number of smaller clusters. On DS3 and DS4, DBSCAN managed to find the genuine clusters with right
parameter values. Figure 12 (f)–(h) shows the sensitivity of DBSCAN with respect to the Eps parameter.

6 Concluding Remarks
In this paper, we have presented a novel hierarchical clustering algorithm called C HAMELEON which takes into account
the dynamic model of clusters. C HAMELEON can discover natural clusters of different shapes and sizes, because its
merging decision dynamically adapts to the different clustering model characterized by the clusters in consideration.
Experimental results on several data sets with varying characteristics show that C HAMELEON can discover natural
clusters that many existing clustering algorithms fail to find.
All the data sets presented in the paper are in 2D space, partly because similar data sets have been used most ex-
tensively by other authors [NH94, EKSX96, GRS98], and partly because it is easy to evaluate the quality of clustering
on 2D data sets. Note that many of these schemes [EKSX96, GRS98] are specifically suited for spatial data and/or
data in metric spaces. Hence, it is noteworthy that our scheme outperforms them, even though it does not make use
of the metric-space/spatial nature of the data. The methodology of dynamic modeling of clusters in agglomerative
hierarchical methods is applicable to all types of data as long as a similarity matrix is available or can be constructed.
Even though we chose to model the data using k-nearest neighbor graph in this paper, it is entirely possible to
use other graph representations suitable for particular application domains, e.g., such as those based upon mutual
shared neighbors [GK78, JD88, GRS99]. Furthermore, different domains may require different models for capturing
relative closeness and inter-connectivity of pairs of clusters. In any of these situations, we believe that the two-
phase framework of C HAMELEON would still be highly effective. Our future research includes the verification of
C HAMELEON on different application domains and the study of effectiveness of different techniques for modeling
data as well as cluster similarity.
In this paper, we ignored the issue of scaling to large data sets that cannot fit in the main memory. These issues are
orthogonal to the ones discussed here and are covered in [ZRL96, BFR98, GRS98, GRG + 99].

15
Acknowledgements
We would like to thank M. Ester, H.P. Kriegel, J. Sander, and X. Xu for providing the DBSCAN program.

References
[Alp98] C. J. Alpert. The ISPD98 circuit benchmark suite. In Proc. of the Intl. Symposium of Physical Design,
pages 80–85, 1998.
[BBK98] S. Berchtold, C. Bohm, and H.-P. Kriegel. They pyramid-technique: Towards breaking the curse of
dimensionality. In Proc. of 1998 ACM-SIGMOD Int. Conf. on Management of Data, 1998.
[BBKK97] S. Berchtold, C. Bohm, D. Keim, and H.-P. Kriegel. Cost model for nearest neighbor search in high-
dimensional data space. In Proc. of Symposium on Principles of Database Systems, Tucson, Arizona,
1997.
[BFR98] P.S. Bradley, U. Fayyad, and C. Reina. Scaling clustering algorithms to large databases. In Proc. of the
Fourth Int’l Conference on Knowledge Discovery and Data Mining, 1998.
[BGG+ 99a] D. Boley, M. Gini, R. Gross, E.H. Han, K. Hastings, G. Karypis, V. Kumar, B. Mobasher, and J. Moore.
Document categorization and query generation on the world wide web using WebACE. AI Review (ac-
cepted for publication), 1999.
[BGG+ 99b] D. Boley, M. Gini, R. Gross, E.H. Han, K. Hastings, G. Karypis, V. Kumar, B. Mobasher, and J. Moore.
Partitioning-based clustering for web document categorization. Decision Support Systems (accepted for
publication), 1999.
[BH64] G.H. Ball and D.J. Hall. Some fundamental concepts and sysnthesis procedures for pattern recognition
preprocessors. In International Conference on Microwaves, Circuit Theory, and Information Theory,
1964.
[BR98] Simon Byers and Adrian E. Raftery. Nearest neighbor clutter removal for estimating features in spatial
point processes. J. American Statistical Association, 93:577–584, June 1998.
[CHY96] M.S. Chen, J. Han, and P.S. Yu. Data mining: An overview from database perspective. IEEE Transac-
tions on Knowledge and Data Eng., 8(6):866–883, December 1996.
[CS96] P. Cheeseman and J. Stutz. Baysian classification (autoclass): Theory and results. In U.M. Fayyad,
G. Piatetsky-Shapiro, P. Smith, and R. Uthurusamy, editors, Advances in Knowledge Discovery and
Data Mining, pages 153–180. AAAI/MIT Press, 1996.
[EKSX96] M. Ester, H.-P. Kriegel, J. Sander, and X. Xu. A density-based algorithm for discovering clusters in large
spatial databases with noise. In Proc. of the Second Int’l Conference on Knowledge Discovery and Data
Mining, Portland, OR, 1996.
[FBF77] J.H. Friedman, J.L. Bentley, and R.A. Finkel. An algorithm for finding best matches in logarithmic
expected time. ACM Transactions on Mathematical Software, 3:209–226, 1977.
[GK78] K.C. Gowda and G. Krishna. Agglomerative clustering using the concept of mutual nearest neighbor-
hood. Pattern Recognition, 10:105–112, 1978.
[GRG+ 99] V. Ganti, R. Ramakrishnan, J. Gehrke, A. Powell, and J. French. Clustering large datasets in arbitrary
metric spaces. In Proc. of the 15th Int’l Conf. on Data Eng., 1999.

16
[GRS98] Sudipto Guha, Rajeev Rastogi, and Kyuseok Shim. CURE: An efficient clustering algorithm for large
databases. In Proc. of 1998 ACM-SIGMOD Int. Conf. on Management of Data, 1998.
[GRS99] Sudipto Guha, Rajeev Rastogi, and Kyuseok Shim. ROCK: a robust clustering algorithm for categorical
attributes. In Proc. of the 15th Int’l Conf. on Data Eng., 1999.
[HHS92] N. Harris, L. Hunter, and D. States. Mega-classification: Discovering motifs in massive datastreams. In
Proceedings of the Tenth International Conference on Artificial Intelligence (AAAI), 1992.
[HKKM97] E.H. Han, G. Karypis, V. Kumar, and B. Mobasher. Clustering based on association rule hypergraphs.
Technical Report TR-97-019, Department of Computer Science, University of Minnesota, Minneapolis,
1997.
[HKKM98] E.H. Han, G. Karypis, V. Kumar, and B. Mobasher. Hypergraph based clustering in high-dimensional
data sets: A summary of results. Bulletin of the Technical Committee on Data Engineering, 21(1), 1998.
[JD88] A.K. Jain and R. C. Dubes. Algorithms for Clustering Data. Prentice Hall, 1988.
[JP73] R.A. Jarvis and E.A. Patrick. Clustering using a similarity measure based on shared nearest neighbors.
IEEE Transactions on Computers, C-22(11), November 1973.
[KK98a] G. Karypis and V. Kumar. hMETIS 1.5: A hypergraph partitioning package. Technical report, De-
partment of Computer Science, University of Minnesota, 1998. Available on the WWW at URL
http://www.cs.umn.edu/˜metis.
[KK98b] G. Karypis and V. Kumar. METIS 4.0: Unstructured graph partitioning and sparse matrix ordering system.
Technical report, Department of Computer Science, University of Minnesota, 1998. Available on the
WWW at URL http://www.cs.umn.edu/˜metis.
[KK98c] G. Karypis and V. Kumar. Multilevel k-way partitioning scheme for irregular graphs. Journal
of Parallel and Distributed Computing, 48(1):96–129, 1998. Also available on WWW at URL
http://www.cs.umn.edu/˜karypis.
[KK99a] G. Karypis and V. Kumar. A fast and highly quality multilevel scheme for partitioning irregular
graphs. SIAM Journal on Scientific Computing, 20(1), 1999. Also available on WWW at URL
http://www.cs.umn.edu/˜karypis. A short version appears in Intl. Conf. on Parallel Processing 1995.
[KK99b] G. Karypis and V. Kumar. Multilevel k-way hypergraph partitioning. In Proceedings of the Design and
Automation Conference, 1999.
[KR90] L. Kaufman and P.J. Rousseeuw. Finding Groups in Data: an Introduction to Cluster Analysis. John
Wiley & Sons, 1990.
[NH94] R. Ng and J. Han. Efficient and effective clustering method for spatial data mining. In Proc. of the 20th
VLDB Conference, pages 144–155, Santiago, Chile, 1994.
[NRS+ 95] T. Newman, E.F. Retzel, E. Shoop, E. Chi, and C. Somerville. Arabidopsis thaliana expressed sequence
tags: Generation, analysis and dissemination. In Plant Genome III: International Conference on the
Status of Plant Genome Research, San Diego, CA, 1995.
[PAS96] L.J. Hubert P. Arabie and G. De Soete. Clustering and Classification. World Scientific, 1996.
[SAD+ 93] M. Stonebraker, R. Agrawal, U. Dayal, E. J. Neuhold, and A. Reuter. DBMS research at a crossroads:
The vienna update. In Proc. of the 19th VLDB Conference, pages 688–692, Dublin, Ireland, 1993.

17
[Sam90] H. Samet. The Design and Analysis of Spatial Data Structures. Addison-Wesley, 1990.
[SCC+ 95] E. Shoop, E. Chi, J. Carlis, P. Bieganski, J. Riedl, N. Dalton, T. Newman, and E. Retzel. Implementation
and testing of an automated EST processing and analysis system. In Lawrence Hunter and Bruce Shriver,
editors, Proceedings of the 28th Annual Hawaii International Conference on System Sciences, volume 5,
pages 52–61. IEEE Computer Society Press, 1995.
[XEKS98] X. Xu, M. Ester, H.-P. Kriegel, and J. Sander. A distribution-based clustering algorithm for mining large
spatial databases. In Proc. of the 14th Int’l Conf. on Data Eng., 1998.
[ZRL96] T. Zhang, R. Ramakrishnan, and M. Linvy. Birch: an efficient data clustering method for large databases.
In Proc. of 1996 ACM-SIGMOD Int. Conf. on Management of Data, Montreal, Quebec, 1996.

18
 DS1 DS2

DS3 DS4

DS5

Figure 10: The clusters discovered by C HAMELEON for the five data sets.

19
 DS3 (17 Clusters) DS3 (11 Clusters)

DS4 (25 Clusters) DS4 (11 Clusters)

DS5 (25 Clusters) DS5 (8 Clusters)

Figure 11: Clusters of CURE with shrinking factor 0.3 and number of representative points 10.

20
 (a) DS1: Eps=0.5, MinPts=4 (b) DS1: Eps=0.4, MinPts=4

(c) DS2: Eps=5.0, MinPts=4 (d) DS2: Eps=3.5, MinPts=4 (e) DS2: Eps=3.0, MinPts=4

(f) DS4: Eps=5.5, MinPts=4 (g) DS4: Eps=5.9, MinPts=4 (h) DS4: Eps=6.2, MinPts=4

Figure 12: DBSCAN on the DS1, DS2, and DS4 data sets with different values of the Eps parameter.

21

View publication stats