Data Mining

Cluster Analysis: Advanced Concepts

and Algorithms

Lecture Notes for Chapter 8

Introduction to Data Mining

by
Tan, Steinbach, Kumar

Introduction to Data Mining, 2nd Edition

Tan, Steinbach, Karpatne, Kumar

Outline

 Prototype-based
– Fuzzy c-means
– Mixture Model Clustering
– Self-Organizing Maps
 Density-based
– Grid-based clustering
– Subspace clustering: CLIQUE
– Kernel-based: DENCLUE
 Graph-based
– Chameleon
– Jarvis-Patrick
– Shared Nearest Neighbor (SNN)
 Characteristics of Clustering Algorithms

3/31/2021 Introduction to Data Mining, 2nd Edition 2

Tan, Steinbach, Karpatne, Kumar

2
 Hard (Crisp) vs Soft (Fuzzy) Clustering

 Hard (Crisp) vs. Soft (Fuzzy) clustering

– For soft clustering allow point to belong to more than
one cluster
– For K-means, generalize objective function
k
𝑆𝑆𝐸 𝑤 𝑑𝑖𝑠𝑡 𝒙 , 𝒄 w
j 1
ij 1

𝑤 : weight with which object xi belongs to cluster 𝒄𝒋

– To minimize SSE, repeat the following steps:

 Fix 𝒄𝒋 and determine 𝑤 (cluster assignment)
 Fix 𝑤 and recompute 𝒄𝒋
– Hard clustering: 𝑤  {0,1}
3/31/2021 Introduction to Data Mining, 2nd Edition 3
Tan, Steinbach, Karpatne, Kumar

Soft (Fuzzy) Clustering: Estimating Weights

c1 x c2

1 2.5 5

𝑆𝑆𝐸 𝑤 2.5 1 𝑤 5 2.5

2.25𝑤 6.25𝑤

SSE(x) has a minimum value of 2.25 when wx1 = 1, wx2 = 0

3/31/2021 Introduction to Data Mining, 2nd Edition 4

Tan, Steinbach, Karpatne, Kumar

4
 Fuzzy C-means
p: fuzzifier (p > 1)
 Objective function
k
𝑆𝑆𝐸 𝑤 𝑑𝑖𝑠𝑡 𝒙 , 𝒄 w
j 1
ij 1

 𝑤 : weight with which object 𝒙 belongs to cluster 𝒄𝒋

 𝑝: is a power for the weight not a superscript and controls how
“fuzzy” the clustering is

– To minimize objective function, repeat the following:

 Fix 𝒄𝒋 and determine 𝑤
 Fix 𝑤 and recompute 𝒄

– Fuzzy c-means clustering: 𝑤 [0,1]

Bezdek, James C. Pattern recognition with fuzzy objective function algorithms. Kluwer Academic Publishers, 1981.

3/31/2021 Introduction to Data Mining, 2nd Edition 5

Tan, Steinbach, Karpatne, Kumar

Fuzzy C-means
c1 x c2

1 2.5 5

𝑆𝑆𝐸 𝑤 2.5 1 𝑤 5 2.5

2.25𝑤 6.25𝑤

SSE(x) has a minimum value of 1.654 when wx1 = 0.74, wx2 = 0.36

3/31/2021 Introduction to Data Mining, 2nd Edition 6

Tan, Steinbach, Karpatne, Kumar

6
 Fuzzy C-means

 Objective function: p: fuzzifier (p > 1)

k
𝑆𝑆𝐸 𝑤 𝑑𝑖𝑠𝑡 𝒙 , 𝒄 w
j 1
ij 1

 Initialization: choose the weights 𝑤 randomly subject to

k
the constraint that w
j 1
ij 1

 Repeat:
– Update centroids: 𝒄𝒋 𝑤 𝒙 / 𝑤

– Update weights: 𝑤 1/𝑑𝑖𝑠𝑡 𝒙 , 𝒄 / 1/𝑑𝑖𝑠𝑡 𝒙 , 𝒄

3/31/2021 Introduction to Data Mining, 2nd Edition 7

Tan, Steinbach, Karpatne, Kumar

Fuzzy K-means Applied to Sample Data

0.95

0.9

0.85

0.8

0.75

0.7

0.65

0.6

0.55

0.5
maximum
membership
3/31/2021 Introduction to Data Mining, 2nd Edition 8
Tan, Steinbach, Karpatne, Kumar

8
 An Example Application: Image Segmentation

 Modified versions of fuzzy c-means have been

used for image segmentation
– Especially fMRI images (functional magnetic
resonance images)
 References
– Gong, Maoguo, Yan Liang, Jiao Shi, Wenping Ma, and Jingjing Ma. "Fuzzy c-means clustering with local
information and kernel metric for image segmentation." Image Processing, IEEE Transactions on 22, no. 2
(2013): 573-584.

From left to right: original images, fuzzy c-means, EM, BCFCM

– Ahmed, Mohamed N., Sameh M. Yamany, Nevin Mohamed, Aly A. Farag, and Thomas Moriarty. "A modified
fuzzy c-means algorithm for bias field estimation and segmentation of MRI data." Medical Imaging, IEEE
Transactions on 21, no. 3 (2002): 193-199.

3/31/2021 Introduction to Data Mining, 2nd Edition 9

Tan, Steinbach, Karpatne, Kumar

Clustering Using Mixture Models

 Idea is to model the set of data points as arising from a

mixture of distributions
– Typically, normal (Gaussian) distribution is used
– But other distributions have been very profitably used

 Clusters are found by estimating the parameters of the

statistical distributions using the Expectation-
Maximization (EM) algorithm
 k-means is a special case of this approach
Provides a compact representation of clusters
The probabilities with which point belongs to each cluster provide a
functionality similar to fuzzy clustering.

3/31/2021 Introduction to Data Mining, 2nd Edition 10

Tan, Steinbach, Karpatne, Kumar

10
 Probabilistic Clustering: Example

 Informal example: consider

modeling the points that
generate the following
histogram.

 Looks like a combination of two

normal (Gaussian) distributions
 Suppose we can estimate the
mean and standard deviation of each normal distribution.
– This completely describes the two clusters
– We can compute the probabilities with which each point belongs to each
cluster
– Can assign each point to the
cluster (distribution) for which it
is most probable.

3/31/2021 Introduction to Data Mining, 2nd Edition 11

Tan, Steinbach, Karpatne, Kumar

11

Probabilistic Clustering: EM Algorithm

Initialize the parameters

Repeat
For each point, compute its probability under each distribution
Using these probabilities, update the parameters of each distribution
Until there is no change

 Very similar to of K-means

 Consists of assignment and update steps
 Can use random initialization
– Problem of local minima
 For normal distributions, typically use K-means to initialize
 If using normal distributions, can find elliptical as well as
spherical shapes.

3/31/2021 Introduction to Data Mining, 2nd Edition 12

Tan, Steinbach, Karpatne, Kumar

12
 Probabilistic Clustering: Updating Centroids

Update formula for weights assuming an estimate for statistical parameters

m m
c j   xi p (C j | xi ) /  p (C j | xi )
𝒙𝒊 𝑖𝑠 𝑎 𝑑𝑎𝑡𝑎 𝑝𝑜𝑖𝑛𝑡
𝐶𝒋 𝑖𝑠 𝑎 𝑐𝑙𝑢𝑠𝑡𝑒𝑟
𝒄𝒋 𝑖𝑠 𝑎 𝑐𝑒𝑛𝑡𝑟𝑜𝑖𝑑
i 1 i 1

 Very similar to the fuzzy k-means formula

– Weights are probabilities
– Weights are not raised to a power 𝑝 𝒙𝒊 𝐶𝒋 𝑝 𝐶𝒋
– Probabilities calculated using Bayes rule: 𝑝 𝐶𝒋 𝒙𝒊 ∑ 𝑝 𝒙𝒊 𝐶 𝑝 𝐶𝒍
 Need to assign weights to each cluster
– Weights may not be equal
– Similar to prior probabilities
1 m
– Can be estimated: p (C j )   p(C j | xi )
m i 1
3/31/2021 Introduction to Data Mining, 2nd Edition 13
Tan, Steinbach, Karpatne, Kumar

13

More Detailed EM Algorithm

3/31/2021 Introduction to Data Mining, 2nd Edition 14

Tan, Steinbach, Karpatne, Kumar

14
 Probabilistic Clustering Applied to Sample Data

0.95

0.9

0.85

0.8

0.75

0.7

0.65

0.6

0.55

0.5
maximum
probability
3/31/2021 Introduction to Data Mining, 2nd Edition 15
Tan, Steinbach, Karpatne, Kumar

15

Probabilistic Clustering: Dense and Sparse Clusters

6

?
4

0
y

-2

-4

-6

-8
-10 -8 -6 -4 -2 0 2 4
x

3/31/2021 Introduction to Data Mining, 2nd Edition 16

Tan, Steinbach, Karpatne, Kumar

16
 Problems with EM

 Convergence can be slow

 Only guarantees finding local maxima

 Makes some significant statistical assumptions

 Number of parameters for Gaussian distribution

grows as O(d2), d the number of dimensions
– Parameters associated with covariance matrix
– K-means only estimates cluster means, which grow
as O(d)
3/31/2021 Introduction to Data Mining, 2nd Edition 17
Tan, Steinbach, Karpatne, Kumar

17

Alternatives to EM

 Method of moments / Spectral methods

– ICML 2014 workshop bibliography

https://sites.google.com/site/momentsicml2014/bibliography

 Markov chain Monte Carlo (MCMC)

 Other approaches

3/31/2021 Introduction to Data Mining, 2nd Edition 18

Tan, Steinbach, Karpatne, Kumar

18
 SOM: Self-Organizing Maps

 Self-organizing maps (SOM)

– Centroid based clustering scheme
– Like K-means, a fixed number of clusters are specified
– However, the spatial relationship
of clusters is also specified,
typically as a grid
– Points are considered one by one
– Each point is assigned to the
closest centroid, and this centroid is updated
– Other centroids are updated based
on their spatial proximity to the closest centroid
Kohonen, Teuvo, and Self-Organizing Maps. "Springer series in information sciences." Self-
organizing maps 30 (1995).

3/31/2021 Introduction to Data Mining, 2nd Edition 19

Tan, Steinbach, Karpatne, Kumar

19

SOM: Self-Organizing Maps

 Updates are weighted by distance

– Centroids farther away are affected less
 The impact of the updates decreases with each time
– At some point the centroids will not change much
 SOM can be viewed as a type of dimensionality reduction
– If a 2D (3D) grid is used, the results can be easily visualized, and it can facilitate the
interpretation of clusters

3/31/2021 Introduction to Data Mining, 2nd Edition 20

Tan, Steinbach, Karpatne, Kumar

20
 SOM Clusters of LA Times Document Data

3/31/2021 Introduction to Data Mining, 2nd Edition 21

Tan, Steinbach, Karpatne, Kumar

21

Another SOM Example: 2D Points

3/31/2021 Introduction to Data Mining, 2nd Edition 22

Tan, Steinbach, Karpatne, Kumar

22
 Issues with SOM

 High computational complexity

 No guarantee of convergence
 Choice of grid and other parameters is somewhat
arbitrary
 Lack of a specific objective function

3/31/2021 Introduction to Data Mining, 2nd Edition 23

Tan, Steinbach, Karpatne, Kumar

23

Grid-based Clustering

 A type of density-based clustering

3/31/2021 Introduction to Data Mining, 2nd Edition 24

Tan, Steinbach, Karpatne, Kumar

24
 Subspace Clustering

 Until now, we found clusters by considering all of the

attributes
 Some clusters may involve only a subset of attributes, i.e.,
subspaces of the data

3/31/2021 Introduction to Data Mining, 2nd Edition 25

Tan, Steinbach, Karpatne, Kumar

25

Clusters in subspaces

3/31/2021 Introduction to Data Mining, 2nd Edition 26

Tan, Steinbach, Karpatne, Kumar

26
 Clusters in subspaces

3/31/2021 Introduction to Data Mining, 2nd Edition 27

Tan, Steinbach, Karpatne, Kumar

27

Clusters in subspaces

3/31/2021 Introduction to Data Mining, 2nd Edition 28

Tan, Steinbach, Karpatne, Kumar

28
 Clique – A Subspace Clustering Algorithm

 A grid-based clustering algorithm that

methodically finds subspace clusters
– Partitions the data space into rectangular units of
equal volume in all possible subspaces
– Measures the density of each unit by the fraction of
points it contains
– A unit is dense if the fraction of overall points it
contains is above a user specified threshold, 
– A cluster is a set of contiguous (touching) dense units

3/31/2021 Introduction to Data Mining, 2nd Edition 29

Tan, Steinbach, Karpatne, Kumar

29

Clique Algorithm

 It is impractical to check volume units in all

possible subspaces, since there is an
exponential number of such units

 Monotone property of density-based clusters:

– If a set of points cannot form a density based cluster
in k dimensions, then the same set of points cannot
form a density based cluster in all possible supersets
of those dimensions

 Very similar to Apriori algorithm

 Can find overlapping clusters

3/31/2021 Introduction to Data Mining, 2nd Edition 30
Tan, Steinbach, Karpatne, Kumar

30
 Clique Algorithm

3/31/2021 Introduction to Data Mining, 2nd Edition 31

Tan, Steinbach, Karpatne, Kumar

31

Limitations of Clique

 Time complexity is exponential in number of

dimensions
– Especially if “too many” dense units are generated at
lower stages

 May fail if clusters are of widely differing

densities, since the threshold is fixed
– Determining appropriate threshold and unit interval
length can be challenging

3/31/2021 Introduction to Data Mining, 2nd Edition 32

Tan, Steinbach, Karpatne, Kumar

32
 Denclue (DENsity CLUstering)

 Based on the notion of kernel-density estimation

– Contribution of each point to the density is given by
an influence or kernel function

Formula and plot of Gaussian Kernel

– Overall density is the sum of the contributions of all
points

3/31/2021 Introduction to Data Mining, 2nd Edition 33

Tan, Steinbach, Karpatne, Kumar

33

Example of Density from Gaussian Kernel

3/31/2021 Introduction to Data Mining, 2nd Edition 34

Tan, Steinbach, Karpatne, Kumar

34
 DENCLUE Algorithm
 Find the density function
 Identify local maxima (density attractors)
 Assign each point to the density attractor
– Follow direction of maximum increase in density

 Define clusters as groups consisting of points associated with density attractor

 Discard clusters whose density attractor has a density less than a user specified
minimum, 
 Combine clusters connected by paths of points that are connected by points with
density above 

3/31/2021 Introduction to Data Mining, 2nd Edition 35

Tan, Steinbach, Karpatne, Kumar

35

Graph-Based Clustering: General Concepts

 Graph-Based clustering needs a proximity graph

– Each edge between two nodes has a weight which is the
proximity between the two points

 Many hierarchical clustering algorithms can be viewed in

graph terms
– MIN (single link) merges subgraph pairs connected with a lowest
weight edge
– Group Average merges subgraph pairs that have high average
connectivity
 In the simplest case, clusters are connected
components in the graph.

3/31/2021 Introduction to Data Mining, 2nd Edition 36

Tan, Steinbach, Karpatne, Kumar

36
 Graph-Based Clustering: Chameleon

 Based on several key ideas

– Sparsification of the proximity graph
– Partitioning the data into clusters that are
relatively pure subclusters of the “true” clusters
– Merging based on preserving characteristics of
clusters

3/31/2021 Introduction to Data Mining, 2nd Edition 37

Tan, Steinbach, Karpatne, Kumar

37

Graph-Based Clustering: Sparsification

 The amount of data that needs to be

processed is drastically reduced
– Sparsification can eliminate more than 99% of the
entries in a proximity matrix
– The amount of time required to cluster the data is
drastically reduced
– The size of the problems that can be handled is
increased

3/31/2021 Introduction to Data Mining, 2nd Edition 38

Tan, Steinbach, Karpatne, Kumar

38
 Graph-Based Clustering: Sparsification …

 Clustering may work better

– Sparsification techniques keep the connections to the most
similar (nearest) neighbors of a point while breaking the
connections to less similar points.
– This reduces the impact of noise and outliers and sharpens
the distinction between clusters.

 Sparsification facilitates the use of graph

partitioning algorithms (or algorithms based
on graph partitioning algorithms)
– Chameleon and Hypergraph-based Clustering

3/31/2021 Introduction to Data Mining, 2nd Edition 39

Tan, Steinbach, Karpatne, Kumar

39

Limitations of Current Merging Schemes

 Existing merging schemes in hierarchical

clustering algorithms are static in nature
– MIN:
 Merge two clusters based on their closeness (or minimum
distance)

– GROUP-AVERAGE:
 Merge two clusters based on their average connectivity

3/31/2021 Introduction to Data Mining, 2nd Edition 40

Tan, Steinbach, Karpatne, Kumar

40
 Limitations of Current Merging Schemes

(a)

(b)

(c)

(d)

Closeness schemes Average connectivity schemes

will merge (a) and (b) will merge (c) and (d)

3/31/2021 Introduction to Data Mining, 2nd Edition 41

Tan, Steinbach, Karpatne, Kumar

41

Chameleon: Clustering Using Dynamic Modeling

 Adapt to the characteristics of the data set to find the

natural clusters
 Use a dynamic model to measure the similarity between
clusters
– Main properties are the relative closeness and relative inter-
connectivity of the cluster
– Two clusters are combined if the resulting cluster shares certain
properties with the constituent clusters
– The merging scheme preserves self-similarity

3/31/2021 Introduction to Data Mining, 2nd Edition 42

Tan, Steinbach, Karpatne, Kumar

42
 Relative Interconnectivity

3/31/2021 Introduction to Data Mining, 2nd Edition 43

Tan, Steinbach, Karpatne, Kumar

43

Relative Closeness

3/31/2021 Introduction to Data Mining, 2nd Edition 44

Tan, Steinbach, Karpatne, Kumar

44
 Chameleon: Steps

 Preprocessing Step:
Represent the data by a Graph
– Given a set of points, construct the k-nearest-neighbor (k-NN)
graph to capture the relationship between a point and its k
nearest neighbors
– Concept of neighborhood is captured dynamically (even if region
is sparse)

 Phase 1: Use a multilevel graph partitioning

algorithm on the graph to find a large number of
clusters of well-connected vertices
– Each cluster should contain mostly points from one “true” cluster,
i.e., be a sub-cluster of a “real” cluster

3/31/2021 Introduction to Data Mining, 2nd Edition 45

Tan, Steinbach, Karpatne, Kumar

45

Chameleon: Steps …

 Phase 2: Use Hierarchical Agglomerative

Clustering to merge sub-clusters
– Two clusters are combined if the resulting cluster shares certain
properties with the constituent clusters

– Two key properties used to model cluster similarity:

 Relative Interconnectivity: Absolute interconnectivity of two
clusters normalized by the internal connectivity of the
clusters

 Relative Closeness: Absolute closeness of two clusters

normalized by the internal closeness of the clusters

3/31/2021 Introduction to Data Mining, 2nd Edition 46

Tan, Steinbach, Karpatne, Kumar

46
 Experimental Results: CHAMELEON

3/31/2021 Introduction to Data Mining, 2nd Edition 47

Tan, Steinbach, Karpatne, Kumar

47

Experimental Results: CURE (10 clusters)

3/31/2021 Introduction to Data Mining, 2nd Edition 48

Tan, Steinbach, Karpatne, Kumar

48
 Experimental Results: CURE (15 clusters)

3/31/2021 Introduction to Data Mining, 2nd Edition 49

Tan, Steinbach, Karpatne, Kumar

49

Experimental Results: CHAMELEON

3/31/2021 Introduction to Data Mining, 2nd Edition 50

Tan, Steinbach, Karpatne, Kumar

50
 Experimental Results: CURE (9 clusters)

3/31/2021 Introduction to Data Mining, 2nd Edition 51

Tan, Steinbach, Karpatne, Kumar

51

Experimental Results: CURE (15 clusters)

3/31/2021 Introduction to Data Mining, 2nd Edition 52

Tan, Steinbach, Karpatne, Kumar

52
 Experimental Results: CHAMELEON

3/31/2021 Introduction to Data Mining, 2nd Edition 53

Tan, Steinbach, Karpatne, Kumar

53

Spectral Clustering

 Spectral clustering is a graph-based clustering

approach
– Does a graph partitioning of the proximity graph of a data
set
– Breaks the graph into components, such that
 The nodes in a component are strongly connected to other
nodes in the component
 The nodes in a component are weakly connected to nodes in
other components
 See simple example below (W is the proximity matrix)

3/31/2021 Introduction to Data Mining, 2nd Edition 54

Tan, Steinbach, Karpatne, Kumar

54
 Spectral Clustering

 For the simple graph below, the proximity matrix

can be written as

– Because the graph consists of two connected

components finding clusters is easy.
– More generally, we need an automated approach
 Must be able to handle graphs where the components are not
completely separate
 Spectral graph partitioning provides such an approach
 Based on eigenvalue decomposition of a slight modification
of the proximity matrix.

3/31/2021 Introduction to Data Mining, 2nd Edition 55

Tan, Steinbach, Karpatne, Kumar

55

Clustering via Spectral Graph Partitioning

 Uses an eigenvalue based approach to do the graph

portioning
– Based on the Laplacian matrix (L) of a graph, which is derived
from the proximity matrix (W)
 W is also known as the weighted adjacency matrix

 Define a diagonal matrix D

𝑊 , 𝑖𝑓 𝑖 𝑗
𝐷
0, 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
 kth diagonal entry of D is the sum of the edges of the kth node
of W
– See example below

3/31/2021 Introduction to Data Mining, 2nd Edition 56

Tan, Steinbach, Karpatne, Kumar

56
 Clustering via Spectral Graph Partitioning …

 Define a matrix called the Laplacian of the graph, L

𝐋 𝐃 𝐖
 L has the following properties:
– It is symmetric
– vTLv ≥ 0, i.e., the matrix is positive semi-definite and all
eigenvalues are positive.
– Last eigenvalue is 0
 Can write L in terms of its eigenvalue decomposition as
– 𝐋 𝚲𝐕, where 𝚲 is a diagonal matrix of the eigenvalues and
𝐕 is the matrix of eigenvectors

3/31/2021 Introduction to Data Mining, 2nd Edition 57

Tan, Steinbach, Karpatne, Kumar

57

A Slightly More Complicated Example

3/31/2021 Introduction to Data Mining, 2nd Edition 58

Tan, Steinbach, Karpatne, Kumar

58
 Spectral Graph Clustering Algorithm

 Given the Laplacian of a graph, it is easy to

define a spectral graph clustering algorithm
 We simply apply k-means to the matrix
consisting of the first k eignenvectors of L
 Note that we cluster the rows of that matrix

3/31/2021 Introduction to Data Mining, 2nd Edition 59

Tan, Steinbach, Karpatne, Kumar

59

Application of K-means and Spectral Clustering to a 2-D Ring

3/31/2021 Introduction to Data Mining, 2nd Edition 60

Tan, Steinbach, Karpatne, Kumar

60
 Strengths and Limitations

 Can detect clusters of different shape and sizes

 Sensitive to how graph is created and sparsified

 Sensitive to outliers

 Time complexity depends on the sparsity of the

data matrix
– Improved by sparsification

3/31/2021 Introduction to Data Mining, 2nd Edition 61

Tan, Steinbach, Karpatne, Kumar

61

Graph-Based Clustering: SNN Approach

Shared Nearest Neighbor (SNN) graph: the weight of an edge is

the number of shared nearest neighbors between vertices given
that the vertices are connected

i j i j
4

3/31/2021 Introduction to Data Mining, 2nd Edition 62

Tan, Steinbach, Karpatne, Kumar

62
 Graph-Based Clustering: SNN Approach

Shared Nearest Neighbor (SNN) graph: the weight of an edge is

the number of shared neighbors between vertices given that the
vertices are connected

i j i j
4

If two points are similar to many of the same points, then they are
likely similar to one another, even if a direct measurement of
similarity does not indicate this.

3/31/2021 Introduction to Data Mining, 2nd Edition 63

Tan, Steinbach, Karpatne, Kumar

63

Creating the SNN Graph

Sparse Graph Shared Near Neighbor Graph

Link weights are similarities Link weights are number of

between neighboring points Shared Nearest Neighbors

3/31/2021 Introduction to Data Mining, 2nd Edition 64

Tan, Steinbach, Karpatne, Kumar

64
 Jarvis-Patrick Clustering

 First, the k-nearest neighbors of all points are found

– In graph terms this can be regarded as breaking all but the k
strongest links from a point to other points in the proximity graph

 A pair of points is put in the same cluster if

– any two points share more than T neighbors and
– the two points are in each others k nearest neighbor list

 For instance, we might choose a nearest neighbor list of

size 20 and put points in the same cluster if they share
more than 10 near neighbors

 Jarvis-Patrick clustering is too brittle

3/31/2021 Introduction to Data Mining, 2nd Edition 65
Tan, Steinbach, Karpatne, Kumar

65

When Jarvis-Patrick Works Reasonably Well

Original Points Jarvis Patrick Clustering

6 shared neighbors out of 20

3/31/2021 Introduction to Data Mining, 2nd Edition 66

Tan, Steinbach, Karpatne, Kumar

66
 When Jarvis-Patrick Does NOT Work Well

Smallest threshold, T, Threshold of T - 1

that does not merge
clusters.

3/31/2021 Introduction to Data Mining, 2nd Edition 67

Tan, Steinbach, Karpatne, Kumar

67

SNN Density-Based Clustering

 Combines:
– Graph based clustering (similarity definition based on
number of shared nearest neighbors)
– Density based clustering (DBScan-like approach)

 SNN density measures whether a point is

surrounded by similar points (with respect to its
nearest neighbors)

3/31/2021 Introduction to Data Mining, 2nd Edition 68

Tan, Steinbach, Karpatne, Kumar

68
 SNN Clustering Algorithm

1. Compute the similarity matrix

This corresponds to a similarity graph with data points for nodes and
edges whose weights are the similarities between data points
2. Sparsify the similarity matrix by keeping only the k most similar
neighbors
This corresponds to only keeping the k strongest links of the similarity
graph
3. Construct the shared nearest neighbor graph from the sparsified
similarity matrix.
At this point, we could apply a similarity threshold and find the
connected components to obtain the clusters (Jarvis-Patrick
algorithm)
4. Find the SNN density of each Point.
Using a user specified parameters, Eps, find the number points that
have an SNN similarity of Eps or greater to each point. This is the
SNN density of the point
3/31/2021 Introduction to Data Mining, 2nd Edition 69
Tan, Steinbach, Karpatne, Kumar

69

SNN Clustering Algorithm …

5. Find the core points

Using a user specified parameter, MinPts, find the core
points, i.e., all points that have an SNN density greater
than MinPts
6. Form clusters from the core points
If two core points are within a “radius”, Eps, of each
other they are placed in the same cluster
7. Discard all noise points
All non-core points that are not within a “radius” of Eps of
a core point are discarded
8. Assign all non-noise, non-core points to clusters
This can be done by assigning such points to the
nearest core point
(Note that steps 4-8 are DBSCAN)
3/31/2021 Introduction to Data Mining, 2nd Edition 70
Tan, Steinbach, Karpatne, Kumar

70
 SNN Density

a) All Points b) High SNN Density

c) Medium SNN Density d) Low SNN Density

3/31/2021 Introduction to Data Mining, 2nd Edition 71

Tan, Steinbach, Karpatne, Kumar

71

SNN Clustering Can Handle Differing Densities

Original Points SNN Clustering

3/31/2021 Introduction to Data Mining, 2nd Edition 72

Tan, Steinbach, Karpatne, Kumar

72
 SNN Clustering Can Handle Other Difficult Situations

3/31/2021 Introduction to Data Mining, 2nd Edition 73

Tan, Steinbach, Karpatne, Kumar

73

SNN Clusters in Sea-Surface Temperature (SST) Time-series Data

Steinbach, et al (KDD 2003)

107 SNN Clusters for Detrended Monthly Z SST (19 58-1998)

90

60

30
latitude

-30

-60

-90
- 180 -150 -120 -90 -60 -30 0 30 60 90 120 150 180
longitude
3/31/2021 Introduction to Data Mining, 2nd Edition 74
Tan, Steinbach, Karpatne, Kumar

74
 SST Clusters that Correspond to El Nino Climate Indices
Steinbach, et al (KDD 2003)

EL Nino Related SST Clusters

90
Niño Range Range
Region Longitude Latitude
60
1+2 (94) 90°W-80°W 10°S-0°
3 (67) 150°W-90°W 5°S-5°N
30
3.4 (78) 170°W-120°W 5°S-5°N
4 (75) 160°E-150°W 5°S-5°N
latitude

75 78 67 94
0
El Nino Regions Defined
-30 by Earth Scientists

-60

-90
-180 -150 -120 -90 -60 -30 0 30 60 90 120 150 180
longitude

SNN clusters of SST that are highly correlated

with El Nino indices, ~ 0.93 correlation.
3/31/2021 Introduction to Data Mining, 2nd Edition 75
Tan, Steinbach, Karpatne, Kumar

75

SNN Clusters in Sea-Level-Pressure (SLP) Time-series Data

Steinbach, et al (KDD 2003)

SN N Density of S LP T im e Series Data

26 SLP Clusters via Shared Nearest Neighbor Clustering (100 NN, 1982-1994) 90
90

24 22
25
60
60

26 13 14

30
30
21
latitude

16 20 17 18 15
0
latitude

19
23 -30
-3 0

1 9

-60
-6 0 3 4
6 2 5
7
11 12 10 -90
8 -180 -15 0 -12 0 -90 -60 -30 0 30 60 90 1 20 1 50 1 80
-9 0
-180 -150 -120 -90 -60 -30 0 30 60 90 1 20 1 50 1 80 longitude
longi tude

SNN Clusters of SLP. SNN Density of Points on the Globe.

3/31/2021 Introduction to Data Mining, 2nd Edition 76

Tan, Steinbach, Karpatne, Kumar

76
 Pairs of SLP Clusters that Correspond to El-Nino SOI

SLP Clusters 15 and 20 SOI vs. Cluster Centroid 20 - C luster Centroid 15 ( c orr = 0.78 )
3 3
Cluster 15 SOI
Cluster 20 20 - 15

2
2

1
1

-1

-1
-2

-2
-3

-3 -4
82 83 84 85 86 87 88 89 90 91 92 93 94 82 83 84 85 86 87 88 89 90 91 92 93 94

Centroids of SLP clusters 15 and 20 Centroid of cluster 20 – Centroid of cluster 15

(near Darwin, Australia and Tahiti) versus SOI
1982-1993.

3/31/2021 Introduction to Data Mining, 2nd Edition 77

Tan, Steinbach, Karpatne, Kumar

77

Limitations of SNN Clustering

 Complexity of SNN Clustering is high

– O( n * time to find numbers of neighbor within Eps)
– In worst case, this is O(n2)
– For lower dimensions, there are more efficient ways to find
the nearest neighbors
 R* Tree
 k-d Trees

 Parameterization is not easy

3/31/2021 Introduction to Data Mining, 2nd Edition 78

Tan, Steinbach, Karpatne, Kumar

78
 Characteristics of Data, Clusters, and
Clustering Algorithms
 A cluster analysis is affected by characteristics of
– Data
– Clusters
– Clustering algorithms

 Looking at these characteristics gives us a

number of dimensions that you can use to
describe clustering algorithms and the results that
they produce

© Tan,Steinbach, Kumar Introduction to Data Mining 4/13/2006 79

79

Characteristics of Data
 High dimensionality
– Dimensionality reduction
 Types of attributes
– Binary, discrete, continuous, asymmetric
– Mixed attribute types (e.g., some are continuous, others nominal)
 Differences in attribute scales
– Normalization techniques
 Size of data set
 Noise and Outliers
 Properties of the data space
– Can you define a meaningful centroid or a meaningful notion of
density
3/31/2021 Introduction to Data Mining, 2nd Edition 80
Tan, Steinbach, Karpatne, Kumar

80
 Characteristics of Clusters

 Data distribution
– Parametric models

 Shape
– Globular or arbitrary shape

 Differing sizes
 Differing densities
 Level of separation among clusters
 Relationship among clusters
 Subspace clusters

3/31/2021 Introduction to Data Mining, 2nd Edition 81

Tan, Steinbach, Karpatne, Kumar

81

Characteristics of Clustering Algorithms

 Order dependence
 Non-determinism
 Scalability
 Number of parameters

3/31/2021 Introduction to Data Mining, 2nd Edition 82

Tan, Steinbach, Karpatne, Kumar

82
 Which Clustering Algorithm?

 Type of Clustering
– Taxonomy vs flat

 Type of Cluster
– Prototype vs connected reguions vs density-based

 Characteristics of Clusters
– Subspace clusters, spatial inter-relationships

 Characteristics of Data Sets and Attributes

 Noise and Outliers
 Number of Data Objects
 Number of Attributes
 Algorithmic Considerations
3/31/2021 Introduction to Data Mining, 2nd Edition 83
Tan, Steinbach, Karpatne, Kumar

83

Comparison of MIN and EM-Clustering

 We assume EM clustering using the Gaussian (normal) distribution.
 MIN is hierarchical, EM clustering is partitional.
 Both MIN and EM clustering are complete.
 MIN has a graph-based (contiguity-based) notion of a cluster, while
EM clustering has a prototype (or model-based) notion of a cluster.
 MIN will not be able to distinguish poorly separated clusters, but EM
can manage this in many situations.
 MIN can find clusters of different shapes and sizes; EM clustering
prefers globular clusters and can have trouble with clusters of
different sizes.
 Min has trouble with clusters of different densities, while EM can
often handle this.
 Neither MIN nor EM clustering finds subspace clusters.
3/31/2021 Introduction to Data Mining, 2nd Edition 84
Tan, Steinbach, Karpatne, Kumar

84
 Comparison of MIN and EM-Clustering

 MIN can handle outliers, but noise can join clusters; EM clustering
can tolerate noise, but can be strongly affected by outliers.
 EM can only be applied to data for which a centroid is meaningful;
MIN only requires a meaningful definition of proximity.
 EM will have trouble as dimensionality increases and the number
of its parameters (the number of entries in the covariance matrix)
increases as the square of the number of dimensions; MIN can
work well with a suitable definition of proximity.
 EM is designed for Euclidean data, although versions of EM
clustering have been developed for other types of data. MIN is
shielded from the data type by the fact that it uses a similarity
matrix.
 MIN makes no distribution assumptions; the version of EM we are
considering assumes Gaussian distributions.
3/31/2021 Introduction to Data Mining, 2nd Edition 85
Tan, Steinbach, Karpatne, Kumar

85

Comparison of MIN and EM-Clustering

 EM has an O(n) time complexity; MIN is O(n2log(n)).

 Because of random initialization, the clusters found by EM
can vary from one run to another; MIN produces the same
clusters unless there are ties in the similarity matrix.
 Neither MIN nor EM automatically determine the number
of clusters.
 MIN does not have any user-specified parameters; EM
has the number of clusters and possibly the weights of
the clusters.
 EM clustering can be viewed as an optimization problem;
MIN uses a graph model of the data.
 Neither EM or MIN are order dependent.
3/31/2021 Introduction to Data Mining, 2nd Edition 86
Tan, Steinbach, Karpatne, Kumar

86
 Comparison of DBSCAN and K-means

 Both are partitional.

 K-means is complete; DBSCAN is not.
 K-means has a prototype-based notion of a cluster; DB
uses a density-based notion.
 K-means can find clusters that are not well-separated.
DBSCAN will merge clusters that touch.
 DBSCAN handles clusters of different shapes and sizes;
K-means prefers globular clusters.

3/31/2021 Introduction to Data Mining, 2nd Edition 87

Tan, Steinbach, Karpatne, Kumar

87

Comparison of DBSCAN and K-means

 DBSCAN can handle noise and outliers; K-means

performs poorly in the presence of outliers
 K-means can only be applied to data for which a centroid
is meaningful; DBSCAN requires a meaningful definition
of density
 DBSCAN works poorly on high-dimensional data; K-
means works well for some types of high-dimensional
data
 Both techniques were designed for Euclidean data, but
extended to other types of data
 DBSCAN makes no distribution assumptions; K-means is
really assuming spherical Gaussian distributions

3/31/2021 Introduction to Data Mining, 2nd Edition 88

Tan, Steinbach, Karpatne, Kumar

88
 Comparison of DBSCAN and K-means

 K-means has an O(n) time complexity; DBSCAN is

O(n^2)
 Because of random initialization, the clusters found by K-
means can vary from one run to another; DBSCAN
always produces the same clusters
 DBSCAN automatically determines the number of
clusters; K-means does not
 K-means has only one parameter, DBSCAN has two.
 K-means clustering can be viewed as an optimization
problem and as a special case of EM clustering; DBSCAN
is not based on a formal model.

3/31/2021 Introduction to Data Mining, 2nd Edition 89

Tan, Steinbach, Karpatne, Kumar

89