Download Session 9: Clustering

Document related concepts

Principal component analysis wikipedia , lookup

Human genetic clustering wikipedia , lookup

Nonlinear dimensionality reduction wikipedia , lookup

K-nearest neighbors algorithm wikipedia , lookup

K-means clustering wikipedia , lookup

Nearest-neighbor chain algorithm wikipedia , lookup

Cluster analysis wikipedia , lookup

Transcript
Data Mining
Session 9 – Main Theme
Clustering
Dr. Jean-Claude Franchitti
New York University
Computer Science Department
Courant Institute of Mathematical Sciences
Adapted from course textbook resources
Data Mining Concepts and Techniques (2nd Edition)
Jiawei Han and Micheline Kamber
1
Agenda
11
Session
Session Overview
Overview
22
Clustering
Clustering
33
Summary
Summary and
and Conclusion
Conclusion
2
What is the class about?
ƒ Course description and syllabus:
» http://www.nyu.edu/classes/jcf/g22.3033-002/
» http://www.cs.nyu.edu/courses/spring10/G22.3033-002/index.html
ƒ Textbooks:
» Data Mining: Concepts and Techniques (2nd Edition)
Jiawei Han, Micheline Kamber
Morgan Kaufmann
ISBN-10: 1-55860-901-6, ISBN-13: 978-1-55860-901-3, (2006)
» Microsoft SQL Server 2008 Analysis Services Step by Step
Scott Cameron
Microsoft Press
ISBN-10: 0-73562-620-0, ISBN-13: 978-0-73562-620-31 1st Edition (04/15/09)
3
Session Agenda
ƒ What is Cluster Analysis?
ƒ Types of Data in Cluster Analysis
ƒ A Categorization of Major Clustering Methods
ƒ Partitioning Methods
ƒ Hierarchical Methods
ƒ Density-Based Methods
ƒ Grid-Based Methods
ƒ Model-Based Methods
ƒ Clustering High-Dimensional Data
ƒ Constraint-Based Clustering
ƒ Link-based clustering
ƒ Outlier Analysis
ƒ Summary
4
Icons / Metaphors
Information
Common Realization
Knowledge/Competency Pattern
Governance
Alignment
Solution Approach
55
Agenda
11
Session
Session Overview
Overview
22
Clustering
Clustering
33
Summary
Summary and
and Conclusion
Conclusion
6
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
7
What is Cluster Analysis?
ƒ Cluster: A collection of data objects
» similar (or related) to one another within the same group
» dissimilar (or unrelated) to the objects in other groups
ƒ Cluster analysis
» Finding similarities between data according to the
characteristics found in the data and grouping similar
data objects into clusters
ƒ Unsupervised learning: no predefined classes
ƒ Typical applications
» As a stand-alone tool to get insight into data distribution
» As a preprocessing step for other algorithms
8
8
Clustering for Data Understanding and Applications
ƒ Biology: taxonomy of living things: kindom, phylum, class,
order, family, genus and species
ƒ Information retrieval: document clustering
ƒ Land use: Identification of areas of similar land use in an
earth observation database
ƒ Marketing: Help marketers discover distinct groups in their
customer bases, and then use this knowledge to develop
targeted marketing programs
ƒ City-planning: Identifying groups of houses according to
their house type, value, and geographical location
ƒ Earth-quake studies: Observed earth quake epicenters
should be clustered along continent faults
ƒ Climate: understanding earth climate, find patterns of
atmospheric and ocean
ƒ Economic Science: market resarch
9
Clustering: Rich Applications and Multidisciplinary Efforts
ƒ Pattern Recognition
ƒ Spatial Data Analysis
» Create thematic maps in GIS by clustering feature
spaces
» Detect spatial clusters or for other spatial mining tasks
ƒ Image Processing
ƒ Economic Science (especially market research)
ƒ WWW
» Document classification
» Cluster Weblog data to discover groups of similar
access patterns
10
9
Clustering as Preprocessing Tools (Utility)
ƒ Summarization:
» Preprocessing for regression, PCA, classification, and
association analysis
ƒ Compression:
» Image processing: vector quantization
ƒ Finding K-nearest Neighbors
» Localizing search to one or a small number of clusters
1111
Examples of Clustering Applications
ƒ Marketing: Help marketers discover distinct groups in their
customer bases, and then use this knowledge to develop
targeted marketing programs
ƒ Land use: Identification of areas of similar land use in an
earth observation database
ƒ Insurance: Identifying groups of motor insurance policy
holders with a high average claim cost
ƒ City-planning: Identifying groups of houses according to
their house type, value, and geographical location
ƒ Earth-quake studies: Observed earth quake epicenters
should be clustered along continent faults
12
Quality: What Is Good Clustering?
ƒ A good clustering method will produce high
quality clusters
» high intra-class similarity: cohesive within clusters
» low inter-class similarity: distinctive between clusters
ƒ The quality of a clustering result depends on
both the similarity measure used by the method
and its implementation
ƒ The quality of a clustering method is also
measured by its ability to discover some or all of
the hidden patterns
13
Measure the Quality of Clustering
ƒ Dissimilarity/Similarity metric
» Similarity is expressed in terms of a distance function,
typically metric: d(i, j)
» The definitions of distance functions are usually rather
different for interval-scaled, boolean, categorical,
ordinal ratio, and vector variables
» Weights should be associated with different variables
based on applications and data semantics
ƒ Quality of clustering:
» There is usually a separate “quality” function that
measures the “goodness” of a cluster.
» It is hard to define “similar enough” or “good enough”
• The answer is typically highly subjective
14
Distance Measures for Different Kinds of Data
Discussed in Previous Session on Data Preprocessing:
ƒ Numerical (interval)-based:
» Minkowski Distance:
» Special cases: Euclidean (L2-norm), Manhattan (L1-norm)
ƒ Binary variables:
» symmetric vs. asymmetric (Jaccard coeff.)
ƒ
ƒ
ƒ
ƒ
ƒ
Nominal variables: # of mismatches
Ordinal variables: treated like interval-based
Ratio-scaled variables: apply log-transformation first
Vectors: cosine measure
Mixed variables: weighted combinations
15
Requirements of Clustering in Data Mining
ƒ Scalability
ƒ Ability to deal with different types of attributes
ƒ Ability to handle dynamic data
ƒ Discovery of clusters with arbitrary shape
ƒ Minimal requirements for domain knowledge to
determine input parameters
ƒ Able to deal with noise and outliers
ƒ Insensitive to order of input records
ƒ High dimensionality
ƒ Incorporation of user-specified constraints
ƒ Interpretability and usability
16
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
17
Data Structures
ƒ Data matrix
» (two modes)
ƒ Dissimilarity matrix
» (one mode)
 x 11

 ...
x
 i1
 ...
x
 n1
...
x 1f
...
...
...
...
...
...
...
x if
...
...
...
x nf
...
 0
 d(2,1)

 d(3,1 )

 :
 d ( n ,1)
0
d ( 3,2 )
0
:
d ( n ,2 )
:
...
x 1p 

... 
x ip 

... 
x np 







... 0 
18
Type of data in clustering analysis
ƒ Interval-scaled variables
ƒ Binary variables
ƒ Nominal, ordinal, and ratio variables
ƒ Variables of mixed types
19
Interval-valued variables
ƒ Standardize data
» Calculate the mean absolute deviation:
s f = 1n (| x1 f − m f | + | x2 f − m f | +...+ | xnf − m f |)
where
m f = 1n (x1 f + x2 f
+ ... +
xnf )
.
» Calculate the standardized measurement (z-score)
zif =
xif − m f
sf
ƒ Using mean absolute deviation is more robust
than using standard deviation
20
Similarity and Dissimilarity Between Objects
ƒ Distances are normally used to measure the
similarity or dissimilarity between two data
objects
ƒ Some popular ones include: Minkowski
distance:
d (i, j) = q (| x − x |q + | x − x |q +...+ | x − x |q )
i1
j1
i2
j2
ip
jp
where i = (xi1, xi2, …, xip) and j = (xj1, xj2, …, xjp) are
two p-dimensional data objects, and q is a positive
integer
ƒ If q = 1, d is Manhattan distance
d(i, j) =| x − x | +| x − x | +...+| x − x |
i1 j1 i2 j2
ip jp
21
Similarity and Dissimilarity Between Objects (Cont.)
ƒ If q = 2, d is Euclidean distance:
d (i, j) = (| x − x |2 + | x − x |2 +...+ | x − x |2 )
i1
j1
i2
j2
ip
jp
» Properties
• d(i,j) ≥ 0
• d(i,i) = 0
• d(i,j) = d(j,i)
• d(i,j) ≤ d(i,k) + d(k,j)
ƒ Also, one can use weighted distance,
parametric Pearson product moment
correlation, or other disimilarity measures
22
Binary Variables
Object j
ƒ A contingency table for
binary data
Object i
ƒ Distance measure for
1
1
a
ƒ Distance measure for
asymmetric binary variables:
measure for asymmetric
c+d
p
b+c
a+b+c+d
d (i, j ) =
ƒ Jaccard coefficient (similarity
sum
a +b
c
d
0
sum a + c b + d
d (i, j ) =
symmetric binary variables:
0
b
b+c
a+b+c
sim Jaccard (i, j ) =
a
a+b+c
binary variables):
23
Dissimilarity between Binary Variables
ƒ Example
Name
Jack
Mary
Jim
Gender
M
F
M
Fever
Y
Y
Y
Cough
N
N
P
Test-1
P
P
N
Test-2
N
N
N
Test-3
N
P
N
Test-4
N
N
N
» gender is a symmetric attribute
» the remaining attributes are asymmetric binary
» let the values Y and P be set to 1, and the value N be set to 0
0 + 1
= 0 . 33
2 + 0 + 1
1 + 1
= 0 . 67
d ( jack , jim ) =
1 + 1 + 1
1 + 2
d ( jim , mary ) =
= 0 . 75
1 + 1 + 2
d ( jack , mary
) =
24
Nominal Variables
ƒ A generalization of the binary variable in that it
can take more than 2 states, e.g., red, yellow,
blue, green
ƒ Method 1: Simple matching
» m: # of matches, p: total # of variables
d ( i , j ) = p −p m
ƒ Method 2: use a large number of binary variables
» creating a new binary variable for each of the M
nominal states
25
Ordinal Variables
ƒ An ordinal variable can be discrete or continuous
ƒ Order is important, e.g., rank
ƒ Can be treated like interval-scaled
» replace xif by their rank
r if ∈ {1,..., M f }
» map the range of each variable onto [0, 1] by replacing
i-th object in the f-th variable by
z
if
=
r if − 1
M f − 1
» compute the dissimilarity using methods for intervalscaled variables
26
Ratio-Scaled Variables
ƒ Ratio-scaled variable: a positive measurement on
a nonlinear scale, approximately at exponential
scale,
such as AeBt or Ae-Bt
ƒ Methods:
» treat them like interval-scaled variables—not a good
choice! (why?—the scale can be distorted)
» apply logarithmic transformation
yif = log(xif)
» treat them as continuous ordinal data treat their rank as
interval-scaled
27
Variables of Mixed Types
ƒ A database may contain all the six types of
variables
» symmetric binary, asymmetric binary, nominal,
ordinal, interval and ratio
ƒ One may use a weighted formula to combine
their effects
Σ p
δ ( f )d ( f )
d (i, j ) =
f = 1
ij
p
f = 1
ij
( f )
ij
Σ
δ
» f is binary or nominal:
dij(f) = 0 if xif = xjf , or dij(f) = 1 otherwise
» f is interval-based: use the normalized distance
» f is ordinal or ratio-scaled
• compute ranks rif and
r −1
• and treat zif as interval-scaled z if = M − 1
if
f
28
Vector Objects
ƒ Vector objects: keywords in documents,
gene features in micro-arrays, etc.
ƒ Broad applications: information retrieval,
biologic taxonomy, etc.
ƒ Cosine measure
ƒ A variant: Tanimoto coefficient
29
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
30
Major Clustering Approaches (I)
ƒ Partitioning approach:
» Construct various partitions and then evaluate them by some
criterion, e.g., minimizing the sum of square errors
» Typical methods: k-means, k-medoids, CLARANS
ƒ Hierarchical approach:
» Create a hierarchical decomposition of the set of data (or objects)
using some criterion
» Typical methods: Diana, Agnes, BIRCH, ROCK, CAMELEON
ƒ Density-based approach:
» Based on connectivity and density functions
» Typical methods: DBSACN, OPTICS, DenClue
ƒ Grid-based approach:
» based on a multiple-level granularity structure
» Typical methods: STING, WaveCluster, CLIQUE
31
Major Clustering Approaches (II)
ƒ Model-based:
» A model is hypothesized for each of the clusters and tries to find
the best fit of that model to each other
» Typical methods: EM, SOM, COBWEB
ƒ Frequent pattern-based:
» Based on the analysis of frequent patterns
» Typical methods: p-Cluster
ƒ User-guided or constraint-based:
» Clustering by considering user-specified or application-specific
constraints
» Typical methods: COD (obstacles), constrained clustering
ƒ Link-based clustering:
» Objects are often linked together in various ways
» Massive links can be used to cluster objects: SimRank, LinkClus
32
Calculation of Distance between Clusters
ƒ Single link: smallest distance between an element in one cluster
and an element in the other, i.e., dist(Ki, Kj) = min(tip, tjq)
ƒ Complete link: largest distance between an element in one cluster
and an element in the other, i.e., dist(Ki, Kj) = max(tip, tjq)
ƒ Average: avg distance between an element in one cluster and an
element in the other, i.e., dist(Ki, Kj) = avg(tip, tjq)
ƒ Centroid: distance between the centroids of two clusters, i.e.,
dist(Ki, Kj) = dist(Ci, Cj)
ƒ Medoid: distance between the medoids of two clusters, i.e., dist(Ki,
Kj) = dist(Mi, Mj)
» Medoid: one chosen, centrally located object in the cluster
33
Centroid, Radius and Diameter of a Cluster (for numerical data sets)
ƒ Centroid: the “middle” of a cluster
Cm =
ΣiN= 1(t
ip
)
N
ƒ Radius: square root of average distance from any point
of the cluster to its centroid
Σ N (t − cm ) 2
Rm = i =1 ip
N
ƒ Diameter: square root of average mean squared
distance between all pairs of points in the cluster
Σ N Σ N (t −t )2
Dm = i =1 i =1 ip iq
N ( N −1)
34
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
35
Partitioning Algorithms: Basic Concept
ƒ Partitioning method: Construct a partition of a database D
of n objects into a set of k clusters, s.t., min sum of
squared distance
E = Σik=1Σ p∈Ci ( p − mi ) 2
ƒ Given a k, find a partition of k clusters that optimizes the
chosen partitioning criterion
» Global optimal: exhaustively enumerate all partitions
» Heuristic methods: k-means and k-medoids algorithms
» k-means (MacQueen’67): Each cluster is represented by the center
of the cluster
» k-medoids or PAM (Partition around medoids) (Kaufman &
Rousseeuw’87): Each cluster is represented by one of the objects
in the cluster
36
The K-Means Clustering Method
ƒ Given k, the k-means algorithm is
implemented in four steps:
» Partition objects into k nonempty subsets
» Compute seed points as the centroids of the
clusters of the current partition (the centroid is the
center, i.e., mean point, of the cluster)
» Assign each object to the cluster with the nearest
seed point
» Go back to Step 2, stop when no more new
assignment
37
The K-Means Clustering Method
ƒ Example
10
9
8
7
6
10
10
9
9
8
8
7
7
6
6
5
5
5
4
4
3
2
1
0
0
1
2
3
4
5
6
7
8
K=2
Arbitrarily choose K
object as initial
cluster center
9
10
Assign
each
objects
to most
similar
center
Update
the
cluster
means
3
2
1
0
0
1
2
3
4
5
6
7
8
9
10
4
3
2
1
0
0
1
2
3
4
5
6
10
9
9
8
8
7
7
6
6
5
4
3
2
1
0
1
2
3
4
5
6
7
8
8
9
10
reassign
reassign
10
0
7
9
10
Update
the
cluster
means
5
4
3
2
1
0
0
1
2
3
4
5
6
7
8
9
10
38
Comments on the K-Means Method
ƒ Strength: Relatively efficient: O(tkn), where n is # objects, k
is # clusters, and t is # iterations. Normally, k, t << n.
• Comparing: PAM: O(k(n-k)2 ), CLARA: O(ks2 + k(n-k))
ƒ Comment: Often terminates at a local optimum. The global
optimum may be found using techniques such as:
deterministic annealing and genetic algorithms
ƒ Weakness
» Applicable only when mean is defined, then what about categorical
data?
» Need to specify k, the number of clusters, in advance
» Unable to handle noisy data and outliers
» Not suitable to discover clusters with non-convex shapes
39
Variations of the K-Means Method
ƒ A few variants of the k-means which differ in
» Selection of the initial k means
» Dissimilarity calculations
» Strategies to calculate cluster means
ƒ Handling categorical data: k-modes (Huang’98)
» Replacing means of clusters with modes
» Using new dissimilarity measures to deal with categorical objects
» Using a frequency-based method to update modes of clusters
» A mixture of categorical and numerical data: k-prototype method
40
What Is the Problem of the K-Means Method?
ƒ The k-means algorithm is sensitive to outliers !
» Since an object with an extremely large value may substantially
distort the distribution of the data.
ƒ K-Medoids: Instead of taking the mean value of the object
in a cluster as a reference point, medoids can be used,
which is the most centrally located object in a cluster.
10
10
9
9
8
8
7
7
6
6
5
5
4
4
3
3
2
2
1
1
0
0
0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
41
The K-Medoids Clustering Method
ƒ Find representative objects, called medoids, in clusters
ƒ PAM (Partitioning Around Medoids, 1987)
» starts from an initial set of medoids and iteratively replaces one
of the medoids by one of the non-medoids if it improves the total
distance of the resulting clustering
» PAM works effectively for small data sets, but does not scale
well for large data sets
ƒ CLARA (Kaufmann & Rousseeuw, 1990)
ƒ CLARANS (Ng & Han, 1994): Randomized sampling
ƒ Focusing + spatial data structure (Ester et al., 1995)
42
A Typical K-Medoids Algorithm (PAM)
Total Cost = 20
10
10
10
9
9
9
8
8
Arbitrary
choose k
object as
initial
medoids
7
6
5
4
3
2
8
7
6
5
4
3
2
1
1
0
0
0
1
2
3
4
5
6
7
8
9
0
10
1
2
3
4
5
6
7
8
9
10
Assign
each
remainin
g object
to
nearest
medoids
7
6
5
4
3
2
1
0
0
K=2
10
Until no
change
If quality is
improved.
3
4
5
6
7
8
9
10
10
9
Swapping O
and Oramdom
2
Randomly select a
nonmedoid object,Oramdom
Total Cost = 26
Do loop
1
Compute
total cost of
swapping
8
7
6
9
8
7
6
5
5
4
4
3
3
2
2
1
1
0
0
0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
43
PAM (Partitioning Around Medoids) (1987)
ƒ PAM (Kaufman and Rousseeuw, 1987), built in
Splus
ƒ Use real object to represent the cluster
» Select k representative objects arbitrarily
» For each pair of non-selected object h and selected
object i, calculate the total swapping cost TCih
» For each pair of i and h,
• If TCih < 0, i is replaced by h
• Then assign each non-selected object to the most
similar representative object
» repeat steps 2-3 until there is no change
44
PAM Clustering: Finding the Best Cluster Center
ƒ Case 1: p currently belongs to oj. If oj is replaced by
orandom as a representative object and p is the closest to
one of the other representative object oi, then p is
reassigned to oi
45
What Is the Problem with PAM?
ƒ Pam is more robust than k-means in the
presence of noise and outliers because a
medoid is less influenced by outliers or other
extreme values than a mean
ƒ Pam works efficiently for small data sets but
does not scale well for large data sets.
» O(k(n-k)2 ) for each iteration
where n is # of data,k is # of clusters
ÎSampling-based method
CLARA(Clustering LARge Applications)
46
CLARA (Clustering Large Applications) (1990)
ƒ CLARA (Kaufmann and Rousseeuw in 1990)
» Built in statistical analysis packages, such as SPlus
» It draws multiple samples of the data set, applies
PAM on each sample, and gives the best clustering
as the output
ƒ Strength: deals with larger data sets than PAM
ƒ Weakness:
» Efficiency depends on the sample size
» A good clustering based on samples will not
necessarily represent a good clustering of the whole
data set if the sample is biased
47
CLARANS (“Randomized” CLARA) (1994)
ƒ CLARANS (A Clustering Algorithm based on
Randomized Search) (Ng and Han’94)
» Draws sample of neighbors dynamically
» The clustering process can be presented as searching a
graph where every node is a potential solution, that is, a
set of k medoids
» If the local optimum is found, it starts with new randomly
selected node in search for a new local optimum
ƒ Advantages: More efficient and scalable than both
PAM and CLARA
ƒ Further improvement: Focusing techniques and
spatial access structures (Ester et al.’95)
48
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
49
Hierarchical Clustering
ƒ Use distance matrix as clustering criteria. This
method does not require the number of clusters
k as an input, but needs a termination condition
Step 0
a
b
Step 1
Step 2 Step 3 Step 4
ab
abcde
c
cde
d
de
e
Step 4
agglomerative
(AGNES)
Step 3
Step 2 Step 1 Step 0
divisive
(DIANA)
50
AGNES (Agglomerative Nesting)
ƒ Introduced in Kaufmann and Rousseeuw (1990)
ƒ Implemented in statistical packages, e.g., Splus
ƒ Use the Single-Link method and the dissimilarity matrix
ƒ Merge nodes that have the least dissimilarity
ƒ Go on in a non-descending fashion
ƒ Eventually all nodes belong to the same cluster
10
10
10
9
9
9
8
8
8
7
7
7
6
6
6
5
5
5
4
4
4
3
3
3
2
2
1
1
0
2
1
0
0
1
2
3
4
5
6
7
8
9
10
0
0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
51
Dendrogram: Shows How the Clusters are Merged
Decompose data objects into a several levels of nested partitioning
(tree of clusters), called a dendrogram.
A clustering of the data objects is obtained by cutting the
dendrogram at the desired level, then each connected component
forms a cluster.
52
DIANA (Divisive Analysis)
ƒ Introduced in Kaufmann and Rousseeuw (1990)
ƒ Implemented in statistical analysis packages,
e.g., Splus
ƒ Inverse order of AGNES
ƒ Eventually each node forms a cluster on its own
10
10
10
9
9
9
8
8
8
7
7
7
6
6
6
5
5
5
4
4
4
3
3
3
2
2
2
1
1
1
0
0
1
2
3
4
5
6
7
8
9
10
0
0
0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
53
Extensions to Hierarchical Clustering
ƒ Major weakness of agglomerative clustering methods
» Do not scale well: time complexity of at least O(n2), where n is the
number of total objects
» Can never undo what was done previously
ƒ Integration of hierarchical & distance-based clustering
» BIRCH (1996): uses CF-tree and incrementally adjusts the quality
of sub-clusters
» ROCK (1999): clustering categorical data by neighbor and link
analysis
» CHAMELEON (1999): hierarchical clustering using dynamic
modeling
54
BIRCH (Zhang, Ramakrishnan & Livny, SIGMOD’96)
ƒ Birch: Balanced Iterative Reducing and Clustering using
Hierarchies
ƒ Incrementally construct a CF (Clustering Feature) tree, a
hierarchical data structure for multiphase clustering
» Phase 1: scan DB to build an initial in-memory CF tree (a multi-level
compression of the data that tries to preserve the inherent clustering
structure of the data)
» Phase 2: use an arbitrary clustering algorithm to cluster the leaf
nodes of the CF-tree
ƒ Scales linearly: finds a good clustering with a single scan
and improves the quality with a few additional scans
ƒ Weakness: handles only numeric data, and sensitive to the
order of the data record
55
Clustering Feature Vector in BIRCH
Clustering Feature (CF): CF = (N, LS, SS)
N: Number of data points
LS: linear sum of N points:
N
∑ X
i
i =1
SS: square sum of N points
2
N
∑ X
i =1
CF = (5, (16,30),(54,190))
10
9
i
8
7
6
5
4
3
2
1
0
0
1
2
3
4
5
6
7
8
9
10
(3,4)
(2,6)
(4,5)
(4,7)
(3,8)
56
CF-Tree in BIRCH
ƒ Clustering feature:
» Summary of the statistics for a given subcluster: the 0-th, 1st and
2nd moments of the subcluster from the statistical point of view.
» Registers crucial measurements for computing cluster and utilizes
storage efficiently
A CF tree is a height-balanced tree that stores the
clustering features for a hierarchical clustering
» A nonleaf node in a tree has descendants or “children”
» The nonleaf nodes store sums of the CFs of their children
ƒ A CF tree has two parameters
» Branching factor: specify the maximum number of children
» Threshold: max diameter of sub-clusters stored at the leaf nodes
57
The CF Tree Structure
Root
B=7
CF1
CF2 CF3
CF6
L=6
child1
child2 child3
child6
Non-leaf node
CF1
CF2 CF3
CF5
child1
child2 child3
child5
Leaf node
prev CF CF
1
2
CF6 next
Leaf node
prev CF CF
1
2
CF4 next
58
Birch Algorithm
ƒ Cluster Diameter
1
2
∑ ( xi − x j )
n ( n − 1)
ƒ For each point in the input
» Find closest leaf entry
» Add point to leaf entry, Update CF
» If entry diameter > max_diameter
• split leaf, and possibly parents
ƒ Algorithm is O(n)
ƒ Problems
» Sensitive to insertion order of data points
» We fix size of leaf nodes, so clusters my not be natural
» Clusters tend to be spherical given the radius and diameter
measures
59
ROCK: Clustering Categorical Data
ƒ ROCK: RObust Clustering using linKs
» S. Guha, R. Rastogi & K. Shim, ICDE’99
ƒ Major ideas
» Use links to measure similarity/proximity
» Not distance-based
ƒ Algorithm: sampling-based clustering
» Draw random sample
» Cluster with links
» Label data in disk
ƒ Experiments
» Congressional voting, mushroom data
60
Similarity Measure in ROCK
ƒ
ƒ
Traditional measures for categorical data may not work
well, e.g., Jaccard coefficient
Example: Two groups (clusters) of transactions
»
»
ƒ
Jaccard co-efficient may lead to wrong clustering result
»
»
ƒ
C1. <a, b, c, d, e>: {a, b, c}, {a, b, d}, {a, b, e}, {a, c, d}, {a, c, e},
{a, d, e}, {b, c, d}, {b, c, e}, {b, d, e}, {c, d, e}
C2. <a, b, f, g>: {a, b, f}, {a, b, g}, {a, f, g}, {b, f, g}
C1: 0.2 ({a, b, c}, {b, d, e}} to 0.5 ({a, b, c}, {a, b, d})
C1 & C2: could be as high as 0.5 ({a, b, c}, {a, b, f})
Jaccard co-efficient-based similarity function:
S im ( T1 , T2 ) =
»
Ex. Let T1 = {a, b, c}, T2 = {c, d, e}
Sim ( T 1 , T 2 ) =
{c}
{ a , b , c , d , e}
=
T1 ∩ T2
T1 ∪ T2
1
= 0 .2
5
61
Link Measure in ROCK
ƒ Clusters
» C1:<a, b, c, d, e>: {a, b, c}, {a, b, d}, {a, b, e}, {a, c, d}, {a, c, e}, {a, d, e},
{b, c, d}, {b, c, e}, {b, d, e}, {c, d, e}
» C2: <a, b, f, g>: {a, b, f}, {a, b, g}, {a, f, g}, {b, f, g}
ƒ Neighbors
» Two transactions are neighbors if sim(T1,T2) > threshold
» Let T1 = {a, b, c}, T2 = {c, d, e}, T3 = {a, b, f}
• T1 connected to: {a,b,d}, {a,b,e}, {a,c,d}, {a,c,e}, {b,c,d}, {b,c,e},
{a,b,f}, {a,b,g}
• T2 connected to: {a,c,d}, {a,c,e}, {a,d,e}, {b,c,e}, {b,d,e}, {b,c,d}
• T3 connected to: {a,b,c}, {a,b,d}, {a,b,e}, {a,b,g}, {a,f,g}, {b,f,g}
ƒ Link Similarity
» Link similarity between two transactions is the # of common neighbors
» link(T1, T2) = 4, since they have 4 common neighbors
• {a, c, d}, {a, c, e}, {b, c, d}, {b, c, e}
» link(T1, T3) = 3, since they have 3 common neighbors
• {a, b, d}, {a, b, e}, {a, b, g}
62
Rock Algorithm
ƒ Method
» Compute similarity matrix
• Use link similarity
» Run agglomerative hierarchical clustering
» When the data set is big
• Get sample of transactions
• Cluster sample
ƒ Problems:
» Guarantee cluster interconnectivity
• any two transactions in a cluster are very well connected
» Ignores information about closeness of two clusters
• two separate clusters may still be quite connected
63
CHAMELEON: Hierarchical Clustering Using Dynamic Modeling (1999)
ƒ
CHAMELEON: by G. Karypis, E. H. Han, and V. Kumar, 1999
ƒ
Measures the similarity based on a dynamic model
ƒ
»
Two clusters are merged only if the interconnectivity and closeness
(proximity) between two clusters are high relative to the internal
interconnectivity of the clusters and closeness of items within the clusters
»
Cure (Hierarchical clustering with multiple representative objects) ignores
information about interconnectivity of the objects, Rock ignores
information about the closeness of two clusters
A two-phase algorithm
1.
Use a graph partitioning algorithm: cluster objects into a large number of
relatively small sub-clusters
2.
Use an agglomerative hierarchical clustering algorithm: find the genuine
clusters by repeatedly combining these sub-clusters
64
Overall Framework of CHAMELEON
Construct (K-NN)
Partition the Graph
Sparse Graph
Data Set
K-NN Graph
p,q connected if q
among the top k
closest neighbors
of p
Merge Partition
Final Clusters
•Relative interconnectivity:
connectivity of c1,c2 over
internal connectivity
•Relative closeness:
closeness of c1,c2 over
internal closeness
65
CHAMELEON (Clustering Complex Objects)
April 22, 2010
Data Mining: Concepts and Techniques
6666
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
67
Density-Based Clustering Methods
ƒ Clustering based on density (local cluster criterion),
such as density-connected points
ƒ Major features:
» Discover clusters of arbitrary shape
» Handle noise
» One scan
» Need density parameters as termination condition
ƒ Several interesting studies:
» DBSCAN: Ester, et al. (KDD’96)
» OPTICS: Ankerst, et al (SIGMOD’99).
» DENCLUE: Hinneburg & D. Keim (KDD’98)
» CLIQUE: Agrawal, et al. (SIGMOD’98) (more grid-based)
68
Density-Based Clustering: Basic Concepts
ƒ Two parameters:
» Eps: Maximum radius of the neighbourhood
» MinPts: Minimum number of points in an Epsneighbourhood of that point
ƒ NEps(p): {q belongs to D | dist(p,q) <= Eps}
ƒ Directly density-reachable: A point p is directly
density-reachable from a point q w.r.t. Eps, MinPts
if
» p belongs to NEps(q)
» core point condition:
p
MinPts = 5
q
Eps = 1 cm
|NEps (q)| >= MinPts
69
Density-Reachable and Density-Connected
ƒ Density-reachable:
» A point p is density-reachable from
a point q w.r.t. Eps, MinPts if there
is a chain of points p1, …, pn, p1 =
q, pn = p such that pi+1 is directly
density-reachable from pi
p
p1
q
ƒ Density-connected
» A point p is density-connected to a
point q w.r.t. Eps, MinPts if there is
a point o such that both, p and q
are density-reachable from o w.r.t.
Eps and MinPts
p
q
o
70
DBSCAN: Density Based Spatial Clustering of Applications with Noise
ƒ Relies on a density-based notion of cluster: A
cluster is defined as a maximal set of densityconnected points
ƒ Discovers clusters of arbitrary shape in spatial
databases with noise
Outlier
Border
Core
Eps = 1cm
MinPts = 5
71
DBSCAN: The Algorithm
ƒ Arbitrary select a point p
ƒ Retrieve all points density-reachable from p w.r.t.
Eps and MinPts.
ƒ If p is a core point, a cluster is formed.
ƒ If p is a border point, no points are densityreachable from p and DBSCAN visits the next
point of the database.
ƒ Continue the process until all of the points have
been processed.
72
DBSCAN: Sensitive to Parameters
73
CHAMELEON (Clustering Complex Objects)
April 22, 2010
Data Mining: Concepts and Techniques
7474
OPTICS: A Cluster-Ordering Method (1999)
ƒ OPTICS: Ordering Points To Identify the
Clustering Structure
» Ankerst, Breunig, Kriegel, and Sander (SIGMOD’99)
» Produces a special order of the database wrt its
density-based clustering structure
» This cluster-ordering contains info equiv to the densitybased clusterings corresponding to a broad range of
parameter settings
» Good for both automatic and interactive cluster analysis,
including finding intrinsic clustering structure
» Can be represented graphically or using visualization
techniques
75
OPTICS: Some Extension from DBSCAN
ƒ Index-based:
•
•
•
•
k = number of dimensions
N = 20
p = 75%
M = N(1-p) = 5
D
» Complexity: O(NlogN)
ƒ Core Distance:
» min eps s.t. point is core
p1
ƒ Reachability Distance
o
p2
Max (core-distance (o), d (o, p))
r(p1, o) = 2.8cm. r(p2,o) = 4cm
o
MinPts = 5
ε = 3 cm
76
Reachability
-distance
undefined
ε
ε‘
ε
Cluster-order
of the objects
77
Density-Based Clustering: OPTICS & Its Applications
78
DENCLUE: Using Statistical Density Functions
ƒ DENsity-based CLUstEring by Hinneburg & Keim
(KDD’98)
ƒ Using statistical density functions:
d ( x, y)2
−
2σ2
f
D
Gaussian
∇f
D
Gaussian
fGaussian (x, y) = e
influence of y
on x
ƒ Major features
(x) =
∑
N
i =1
e
−
total influence
on x
d ( x , xi ) 2
2σ
2
( x, xi ) = ∑i =1 ( xi − x) ⋅ e
N
» Solid mathematical foundation
−
d ( x , xi ) 2
2σ 2
gradient of x in
the direction of
xi
» Good for data sets with large amounts of noise
» Allows a compact mathematical description of arbitrarily shaped
clusters in high-dimensional data sets
» Significant faster than existing algorithm (e.g., DBSCAN)
» But needs a large number of parameters
79
Denclue: Technical Essence
ƒ Uses grid cells but only keeps information about grid cells that do
actually contain data points and manages these cells in a treebased access structure
ƒ Influence function: describes the impact of a data point within its
neighborhood
ƒ Overall density of the data space can be calculated as the sum of
the influence function of all data points
ƒ Clusters can be determined mathematically by identifying density
attractors
ƒ Density attractors are local maximal of the overall density function
ƒ Center defined clusters: assign to each density attractor the points
density attracted to it
ƒ Arbitrary shaped cluster: merge density attractors that are
connected through paths of high density (> threshold)
80
Density Attractor
81
Center-Defined and Arbitrary
82
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
83
Grid-Based Clustering Method
ƒ Using multi-resolution grid data structure
ƒ Several interesting methods
» STING (a STatistical INformation Grid approach) by
Wang, Yang and Muntz (1997)
» WaveCluster by Sheikholeslami, Chatterjee, and
Zhang (VLDB’98)
• A multi-resolution clustering approach using
wavelet method
» CLIQUE: Agrawal, et al. (SIGMOD’98)
• On high-dimensional data (thus put in the
section of clustering high-dimensional data
84
STING: A Statistical Information Grid Approach
ƒ Wang, Yang and Muntz (VLDB’97)
ƒ The spatial area area is divided into rectangular
cells
ƒ There are several levels of cells corresponding to
different levels of resolution
85
The STING Clustering Method
ƒ Each cell at a high level is partitioned into a number of
smaller cells in the next lower level
ƒ Statistical info of each cell is calculated and stored
beforehand and is used to answer queries
ƒ Parameters of higher level cells can be easily calculated
from parameters of lower level cell
» count, mean, s, min, max
» type of distribution—normal, uniform, etc.
ƒ Use a top-down approach to answer spatial data queries
ƒ Start from a pre-selected layer—typically with a small
number of cells
ƒ For each cell in the current level compute the confidence
interval
86
STING Algorithm and Its Analysis
ƒ Remove the irrelevant cells from further
consideration
ƒ When finish examining the current layer,
proceed to the next lower level
ƒ Repeat this process until the bottom layer is
reached
ƒ Advantages:
» Query-independent, easy to parallelize, incremental
update
» O(K), where K is the number of grid cells at the
lowest level
ƒ Disadvantages:
» All the cluster boundaries are either horizontal or
vertical, and no diagonal boundary is detected
87
WaveCluster: Clustering by Wavelet Analysis (1998)
ƒ Sheikholeslami, Chatterjee, and Zhang (VLDB’98)
ƒ A multi-resolution clustering approach which applies
wavelet transform to the feature space
ƒ How to apply wavelet transform to find clusters
» Summarizes the data by imposing a multidimensional grid
structure onto data space
» These multidimensional spatial data objects are represented in a
n-dimensional feature space
» Apply wavelet transform on feature space to find the dense
regions in the feature space
» Apply wavelet transform multiple times which result in clusters at
different scales from fine to coarse
88
Wavelet Transform
ƒ Wavelet transform: A signal processing technique that
decomposes a signal into different frequency sub-band
(can be applied to n-dimensional signals)
ƒ Data are transformed to preserve relative distance
between objects at different levels of resolution
ƒ Allows natural clusters to become more distinguishable
89
The WaveCluster Algorithm
ƒ Input parameters
» # of grid cells for each dimension
» the wavelet, and the # of applications of wavelet transform
ƒ Why is wavelet transformation useful for clustering?
» Use hat-shape filters to emphasize region where points cluster,
but simultaneously suppress weaker information in their boundary
» Effective removal of outliers, multi-resolution, cost effective
ƒ Major features:
»
»
»
»
Complexity O(N)
Detect arbitrary shaped clusters at different scales
Not sensitive to noise, not sensitive to input order
Only applicable to low dimensional data
ƒ Both grid-based and density-based
90
Quantization & Transformation
ƒ First, quantize data into m-D
grid structure, then wavelet
transform
» a) scale 1: high resolution
» b) scale 2: medium resolution
» c) scale 3: low resolution
April 22, 2010
Data Mining: Concepts and Techniques
9191
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
92
Model-Based Clustering
ƒ What is model-based clustering?
» Attempt to optimize the fit between the given data and
some mathematical model
» Based on the assumption: Data are generated by a
mixture of underlying probability distribution
ƒ Typical methods
» Statistical approach
• EM (Expectation maximization), AutoClass
» Machine learning approach
• COBWEB, CLASSIT
» Neural network approach
• SOM (Self-Organizing Feature Map)
93
EM — Expectation Maximization
ƒ EM — A popular iterative refinement algorithm
ƒ An extension to k-means
» Assign each object to a cluster according to a weight (prob.
distribution)
» New means are computed based on weighted measures
ƒ General idea
» Starts with an initial estimate of the parameter vector
» Iteratively rescores the patterns against the mixture density
produced by the parameter vector
» The rescored patterns are used to update the parameter updates
» Patterns belonging to the same cluster, if they are placed by their
scores in a particular component
ƒ Algorithm converges fast but may not be in global optima
94
The EM (Expectation Maximization) Algorithm
ƒ Initially, randomly assign k cluster centers
ƒ Iteratively refine the clusters based on two steps
» Expectation step: assign each data point Xi to cluster Ci
with the following probability
» Maximization step:
• Estimation of model parameters
95
Conceptual Clustering
ƒ Conceptual clustering
» A form of clustering in machine learning
» Produces a classification scheme for a set of unlabeled
objects
» Finds characteristic description for each concept (class)
ƒ COBWEB (Fisher’87)
» A popular a simple method of incremental conceptual
learning
» Creates a hierarchical clustering in the form of a
classification tree
» Each node refers to a concept and contains a
probabilistic description of that concept
96
COBWEB Clustering Method
A classification tree
97
More on Conceptual Clustering
ƒ Limitations of COBWEB
» The assumption that the attributes are independent of each other is
often too strong because correlation may exist
» Not suitable for clustering large database data – skewed tree and
expensive probability distributions
ƒ CLASSIT
» an extension of COBWEB for incremental clustering of continuous
data
» suffers similar problems as COBWEB
ƒ AutoClass (Cheeseman and Stutz, 1996)
» Uses Bayesian statistical analysis to estimate the number of
clusters
» Popular in industry
98
Neural Network Approach
ƒ Neural network approaches
» Represent each cluster as an exemplar, acting as a
“prototype” of the cluster
» New objects are distributed to the cluster whose
exemplar is the most similar according to some
distance measure
ƒ Typical methods
» SOM (Soft-Organizing feature Map)
» Competitive learning
• Involves a hierarchical architecture of several units
(neurons)
• Neurons compete in a “winner-takes-all” fashion for
the object currently being presented
99
Self-Organizing Feature Map (SOM)
ƒ SOMs, also called topological ordered maps, or Kohonen SelfOrganizing Feature Map (KSOMs)
ƒ It maps all the points in a high-dimensional source space into a 2 to 3-d
target space, s.t., the distance and proximity relationship (i.e., topology)
are preserved as much as possible
ƒ Similar to k-means: cluster centers tend to lie in a low-dimensional
manifold in the feature space
ƒ Clustering is performed by having several units competing for the
current object
» The unit whose weight vector is closest to the current object wins
» The winner and its neighbors learn by having their weights adjusted
ƒ SOMs are believed to resemble processing that can occur in the brain
ƒ Useful for visualizing high-dimensional data in 2- or 3-D space
100
Web Document Clustering Using SOM
ƒ The result of
SOM clustering
of 12088 Web
articles
ƒ The picture on
the right: drilling
down on the
keyword
“mining”
ƒ Based on
websom.hut.fi
Web page
101
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
102
Clustering High-Dimensional Data
ƒ Clustering high-dimensional data
» Many applications: text documents, DNA micro-array data
» Major challenges:
• Many irrelevant dimensions may mask clusters
• Distance measure becomes meaningless—due to equi-distance
• Clusters may exist only in some subspaces
ƒ Methods
» Feature transformation: only effective if most dimensions are
relevant
• PCA & SVD useful only when features are highly
correlated/redundant
» Feature selection: wrapper or filter approaches
• useful to find a subspace where the data have nice clusters
» Subspace-clustering: find clusters in all the possible subspaces
• CLIQUE, ProClus, and frequent pattern-based clustering
103
The Curse of Dimensionality
(graphs adapted from Parsons et al. KDD Explorations 2004)
ƒ Data in only one dimension is relatively
packed
ƒ Adding a dimension “stretch” the
points across that dimension, making
them further apart
ƒ Adding more dimensions will make the
points further apart—high dimensional
data is extremely sparse
ƒ Distance measure becomes
meaningless—due to equi-distance
104
Why Subspace Clustering?
(adapted from Parsons et al. SIGKDD Explorations 2004)
ƒ Clusters may exist only in some subspaces
ƒ Subspace-clustering: find clusters in all the
subspaces
April 22, 2010
Data Mining: Concepts and Techniques
105
105
CLIQUE (Clustering In QUEst)
ƒ Agrawal, Gehrke, Gunopulos, Raghavan (SIGMOD’98)
ƒ Automatically identifying subspaces of a high dimensional
data space that allow better clustering than original space
ƒ CLIQUE can be considered as both density-based and gridbased
» It partitions each dimension into the same number of equal length
interval
» It partitions an m-dimensional data space into non-overlapping
rectangular units
» A unit is dense if the fraction of total data points contained in the unit
exceeds the input model parameter
» A cluster is a maximal set of connected dense units within a
subspace
106
CLIQUE: The Major Steps
ƒ Partition the data space and find the number of
points that lie inside each cell of the partition.
ƒ Identify the subspaces that contain clusters using
the Apriori principle
ƒ Identify clusters
» Determine dense units in all subspaces of interests
» Determine connected dense units in all subspaces of
interests.
ƒ Generate minimal description for the clusters
» Determine maximal regions that cover a cluster of
connected dense units for each cluster
» Determination of minimal cover for each cluster
τ=3
30
40
age
60
50
20
30
40
50
age
60
Vacation
20
Vacation
(week)
0 1 2 3 4 5 6 7
Salary
(10,000)
0 1 2 3 4 5 6 7
107
l
Sa
y
ar
30
50
age
108
Strength and Weakness of CLIQUE
ƒ Strength
» automatically finds subspaces of the highest
dimensionality such that high density clusters exist in
those subspaces
» insensitive to the order of records in input and does not
presume some canonical data distribution
» scales linearly with the size of input and has good
scalability as the number of dimensions in the data
increases
ƒ Weakness
» The accuracy of the clustering result may be degraded
at the expense of simplicity of the method
109
Frequent Pattern-Based Approach
ƒ Clustering high-dimensional space (e.g.,
clustering text documents, microarray data)
» Projected subspace-clustering: which dimensions to be
projected on?
• CLIQUE, ProClus
» Feature extraction: costly and may not be effective?
» Using frequent patterns as “features”
ƒ Clustering by pattern similarity in micro-array data
(pClustering) [H. Wang, W. Wang, J. Yang, and
P. S. Yu. Clustering by pattern similarity in large
data sets, SIGMOD’02]
110
Clustering by Pattern Similarity (p-Clustering)
ƒ Right: The micro-array “raw” data
shows 3 genes and their values in a
multi-dimensional space
» Difficult to find their patterns
ƒ Bottom: Some subsets of dimensions
form nice shift and scaling patterns
111
Why p-Clustering?
ƒ
Microarray data analysis may need to
» Clustering on thousands of dimensions (attributes)
» Discovery of both shift and scaling patterns
ƒ
ƒ
Clustering with Euclidean distance measure? — cannot find shift patterns
Clustering on derived attribute Aij = ai – aj? — introduces N(N-1) dimensions
ƒ
Bi-cluster (Y. Cheng and G. Church. Biclustering of expression data. ISMB’00)
using transformed mean-squared residue score matrix (I, J)
» Where
1
d =
∑ d
ij | J |
ij
j∈J
d
Ij
=
1
∑ d
| I | i ∈ I ij
d
IJ
=
1
d
∑
| I || J | i ∈ I , j ∈ J ij
» A submatrix is a δ-cluster if H(I, J) ≤ δ for some δ > 0
ƒ
Problems with bi-cluster
» No downward closure property
» Due to averaging, it may contain outliers but still within δ-threshold
112
p-Clustering:
Clustering by Pattern Similarity
ƒ Given object x, y in O and features a, b in T, pCluster is a 2 by 2
matrix
d xa d xb 
pScore(
) =| (d xa − d xb ) − (d ya − d yb ) |
d ya d yb 
ƒ A pair (O, T) is in δ-pCluster if for any 2 by 2 matrix X in (O, T),
pScore(X) ≤ δ for some δ > 0
ƒ Properties of δ-pCluster
» Downward closure
» Clusters are more homogeneous than bi-cluster (thus the name: pair-wise
Cluster)
ƒ Pattern-growth algorithm has been developed for efficient mining
d
ƒ For scaling patterns, one can observe, taking logarithmic on
d
will lead to the pScore form
xa
/d
ya
xb
/d
yb
< δ
113
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
114
Why Constraint-Based Cluster Analysis?
ƒ
ƒ
Need user feedback: Users know their applications the best
Less parameters but more user-desired constraints, e.g., an
ATM allocation problem: obstacle & desired clusters
115
A Classification of Constraints in Cluster Analysis
ƒ Clustering in applications: desirable to have userguided (i.e., constrained) cluster analysis
ƒ Different constraints in cluster analysis:
» Constraints on individual objects (do selection first)
• Cluster on houses worth over $300K
» Constraints on distance or similarity functions
• Weighted functions, obstacles (e.g., rivers, lakes)
» Constraints on the selection of clustering parameters
• # of clusters, MinPts, etc.
» User-specified constraints
• Contain at least 500 valued customers and 5000 ordinary ones
» Semi-supervised: giving small training sets as
“constraints” or hints
116
Clustering With Obstacle Objects
ƒ Tung, Hou, and Han. Spatial
Clustering in the Presence of
Obstacles, ICDE'01
ƒ K-medoids is more preferable since
k-means may locate the ATM center
in the middle of a lake
ƒ Visibility graph and shortest path
ƒ Triangulation and micro-clustering
ƒ Two kinds of join indices (shortestpaths) worth pre-computation
» VV index: indices for any pair of obstacle
vertices
» MV index: indices for any pair of microcluster and obstacle indices
117
An Example: Clustering With Obstacle Objects
Not Taking obstacles into account
Taking obstacles into account
118
User-Guided Clustering
name
course
course-id
group
office
semester
name
position
instructor
area
Advise
professor
name
student
area
degree
User hint
Publish
Publication
author
title
title
year
conf
Register
student
Target of
clustering
ƒ
ƒ
Course
Professor
person
Group
ƒ
Open-course
Work-In
Student
course
name
office
semester
position
unit
grade
X. Yin, J. Han, P. S. Yu, “Cross-Relational Clustering with User's Guidance”,
KDD'05
User usually has a goal of clustering, e.g., clustering students by research area
User specifies his clustering goal to CrossClus
119
Comparing with Classification
User hint ƒ User-specified feature (in the
form of attribute) is used as a
hint, not class labels
» The attribute may contain too
many or too few distinct values,
e.g., a user may want to
cluster students into 20
clusters instead of 3
All tuples for clustering
» Additional features need to be
included in cluster analysis
120
Comparing with Semi-Supervised Clustering
ƒ Semi-supervised clustering: User provides a training set consisting of
“similar” (“must-link) and “dissimilar” (“cannot link”) pairs of objects
ƒ User-guided clustering: User specifies an attribute as a hint, and
more relevant features are found for clustering
User-guided clustering
All tuples for clustering
Semi-supervised clustering
x
121
Why Not Semi-Supervised Clustering?
ƒ Much information (in multiple relations) is needed to judge
whether two tuples are similar
ƒ A user may not be able to provide a good training set
ƒ It is much easier for a user to specify an attribute as a hint,
such as a student’s research area
Tom Smith
Jane Chang
SC1211
BI205
TA
RA
Tuples to be compared
User hint
122
CrossClus: An Overview
ƒ Measure similarity between features by how they group
objects into clusters
ƒ Use a heuristic method to search for pertinent features
» Start from user-specified feature and gradually expand search
range
ƒ Use tuple ID propagation to create feature values
» Features can be easily created during the expansion of search
range, by propagating IDs
ƒ Explore three clustering algorithms: k-means, kmedoids, and hierarchical clustering
123
Multi-Relational Features
ƒ A multi-relational feature is defined by:
» A join path, e.g., Student → Register → OpenCourse → Course
» An attribute, e.g., Course.area
» (For numerical feature) an aggregation operator, e.g., sum or average
ƒ Categorical feature f = [Student → Register → OpenCourse
→ Course, Course.area, null]
areas of courses of each student
Tuple
t1
t2
t3
t4
t5
Areas of courses
DB
AI
TH
5
5
0
0
3
7
1
5
4
5
0
5
3
3
4
Values of feature f
Tuple
t1
t2
t3
t4
t5
Feature f
DB
AI
TH
0.5
0.5
0
0
0.3 0.7
0.1
0.5 0.4
0.5
0
0.5
0.3
0.3 0.4
f(t1)
f(t2)
f(t3)
f(t4)
DB
AI
TH
f(t5)
124
Representing Features
ƒ Similarity between tuples t1 and t2 w.r.t. categorical feature f
» Cosine similarity between vectors f(t1) and f(t2)
L
sim f (t1 , t 2 ) =
Similarity vector
Vf
∑ f (t ). p
1
k =1
L
∑ f (t ). p
k =1
1
2
k
k
⋅
⋅ f (t 2 ). pk
L
∑ f (t ). p
k =1
2
2
k
ƒ Most important information of a
feature f is how f groups tuples into
clusters
ƒ f is represented by similarities
between every pair of tuples
indicated by f
ƒ The horizontal axes are the tuple
indices, and the vertical axis is the
similarity
ƒ This can be considered as a vector
of N x N dimensions
125
Similarity Between Features
Vf
Values of Feature f and g
Feature f (course)
Feature g (group)
DB
AI
TH
Info sys
Cog sci
Theory
t1
0.5
0.5
0
1
0
0
t2
0
0.3
0.7
0
0
1
t3
0.1
0.5
0.4
0
0.5
0.5
t4
0.5
0
0.5
0.5
0
0.5
t5
0.3
0.3
0.4
0.5
0.5
0
Vg
Similarity between two features –
cosine similarity of two vectors
V f ⋅V g
sim( f , g ) = f g
V V
126
Computing Feature Similarity
Tuples
Feature f
Feature g
DB
Info sys
AI
Cog sci
TH
Theory
Similarity between feature
values w.r.t. the tuples
sim(fk,gq)=Σi=1 to N f(ti).pk∙g(ti).pq
Info sys
DB
2
V ⋅ V = ∑∑ sim f (ti , t j )⋅ simg (ti , t j ) = ∑∑ sim( f k , g q )
f
g
N
N
i =1 j =1
l
k =1 q =1
Tuple similarities,
hard to compute
DB
Info sys
AI
Cog sci
TH
Theory
m
Feature value similarities,
easy to compute
Compute similarity
between each pair of
feature values by one
scan on data
127
Searching for Pertinent Features
ƒ Different features convey different aspects of information
Academic Performances
Research area
Research group area
Conferences of papers
Advisor
Demographic info
GPA
Permanent address
GRE score
Nationality
Number of papers
ƒ Features conveying same aspect of information usually
cluster tuples in more similar ways
» Research group areas vs. conferences of publications
ƒ Given user specified feature
» Find pertinent features by computing feature similarity
128
Heuristic Search for Pertinent Features
Overall procedure
Course
Professor
person
name
course
course-id
group
office
semester
name
position
instructor
area
2
1. Start from the userspecified feature
Group
name
2. Search in
area
neighborhood of
existing pertinent User hint
features
3. Expand search
Target of
range gradually
clustering
„
Open-course
Work-In
Advise
Publish
professor
student
author
1
degree
title
Publication
title
year
conf
Register
student
Student
name
office
position
course
semester
unit
grade
Tuple ID propagation is used to create multi-relational features
„ IDs of target tuples can be propagated along any join path, from
which we can find tuples joinable with each target tuple
129
Clustering with Multi-Relational Features
ƒ Given a set of L pertinent features f1, …, fL,
similarity between two tuples
L
sim(t1 , t2 ) = ∑ sim f i (t1 , t2 ) ⋅ f i .weight
i =1
» Weight of a feature is determined in feature search by
its similarity with other pertinent features
ƒ Clustering methods
» CLARANS [Ng & Han 94], a scalable clustering
algorithm for non-Euclidean space
» K-means
» Agglomerative hierarchical clustering
130
Experiments: Compare CrossClus with
ƒ Baseline: Only use the user specified feature
ƒ PROCLUS [Aggarwal, et al. 99]: a state-of-the-art
subspace clustering algorithm
» Use a subset of features for each cluster
» We convert relational database to a table by
propositionalization
» User-specified feature is forced to be used in every
cluster
ƒ RDBC [Kirsten and Wrobel’00]
» A representative ILP clustering algorithm
» Use neighbor information of objects for clustering
» User-specified feature is forced to be used
131
Measure of Clustering Accuracy
ƒ Accuracy
» Measured by manually labeled data
• We manually assign tuples into clusters according
to their properties (e.g., professors in different
research areas)
» Accuracy of clustering: Percentage of pairs of tuples
in the same cluster that share common label
• This measure favors many small clusters
• We let each approach generate the same number
of clusters
132
DBLP Dataset
Clustering Accurarcy - DBLP
1
0.9
0.8
0.7
CrossClus K-Medoids
CrossClus K-Means
CrossClus Agglm
Baseline
PROCLUS
RDBC
0.6
0.5
0.4
0.3
0.2
0.1
th
re
e
A
ll
ho
r
Co
au
t
W
Co
or
d+
oa
ut
ho
r
or
d
nf
+C
nf
+W
Co
Co
au
th
or
or
d
W
Co
nf
0
133
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
134
Link-Based Clustering: Calculate Similarities Based On Links
Authors
Tom
Mike
Cathy
John
Mary
Proceedings
Conferences
ƒ The similarity between two
objects x and y is defined as
sigmod
the average similarity between
objects linked with x and those
with y:
I ( a ) I (b )
C
vldb
sim(a, b ) =
∑ ∑ sim(I i (a ), I j (b ))
I (a ) I (b ) i =1 j =1
sigmod03
sigmod04
sigmod05
vldb03
vldb04
vldb05
aaai04
aaai05
aaai
ƒ Disadv: Expensive to compute:
Jeh & Widom, KDD’2002: SimRank
Two objects are similar if they are
linked with the same or similar
objects
» For a dataset of N objects and
M links, it takes O(N2) space
and O(M2) time to compute all
similarities.
135
Observation 1: Hierarchical Structures
ƒ Hierarchical structures often exist naturally among
objects (e.g., taxonomy of animals)
Relationships between articles and
words (Chakrabarti, Papadimitriou,
Modha, Faloutsos, 2004)
A hierarchical structure of
products in Walmart
grocery electronics
TV
DVD
apparel
Articles
All
camera
136
Observation 2: Distribution of Similarity
portion of entries
0.4
Distribution of SimRank similarities
among DBLP authors
0.3
0.2
0.1
0.24
0.22
0.2
0.18
0.16
0.14
0.12
0.1
0.08
0.06
0.04
0.02
0
0
similarity value
ƒ Power law distribution exists in similarities
» 56% of similarity entries are in [0.005, 0.015]
» 1.4% of similarity entries are larger than 0.1
» Can we design a data structure that stores the significant
similarities and compresses insignificant ones?
137
A Novel Data Structure: SimTree
Each non-leaf node
represents a group
of similar lower-level
nodes
Each leaf node
represents an object
Similarities between
siblings are stored
Canon A40
digital camera
Digital
Cameras
Sony V3 digital
Consumer
camera
electronics
Apparels
TVs
138
Similarity Defined by SimTree
Similarity between two
sibling nodes n1 and n2
n1
Adjustment ratio
for node n7
0.9
0.8
0.9
n7
n4
ƒ Path-based node similarity
0.3
0.8
n2
0.2
n5
n8
n3
0.9
n6
1.0
n9
» simp(n7,n8) = s(n7, n4) x s(n4, n5) x s(n5, n8)
ƒ Similarity between two nodes is the average similarity
between objects linked with them in other SimTrees
ƒ Adjustment ratio for x =
Average similarity between x and all other nodes
Average similarity between x’s parent and all
other nodes
139
LinkClus: Efficient Clustering via Heterogeneous Semantic Links
X. Yin, J. Han, and P. S. Yu, “LinkClus: Efficient
Clustering via Heterogeneous Semantic Links”,
VLDB'06
Method
ƒ Initialize a SimTree for objects of each type
ƒ Repeat
» For each SimTree, update the similarities between its
nodes using similarities in other SimTrees
• Similarity between two nodes x and y is the average
similarity between objects linked with them
» Adjust the structure of each SimTree
• Assign each node to the parent node that it is most
similar to
140
Initialization of SimTrees
ƒ Initializing a SimTree
» Repeatedly find groups of tightly related nodes, which
are merged into a higher-level node
ƒ Tightness of a group of nodes
» For a group of nodes {n1, …, nk}, its tightness is
defined as the number of leaf nodes in other
SimTrees that are connected to all of {n1, …, nk}
Nodes
n1
n2
Leaf nodes in
another SimTree
1
2
3
4
5
The tightness of {n1, n2} is 3
141
Finding Tight Groups by Freq. Pattern Mining
ƒ Finding tight groups
Frequent pattern mining
Reduced to
The tightness of a
g1
group of nodes is the
support of a frequent
pattern
g2
n1
n2
n3
n4
Transactions
1
2
3
4
5
6
7
8
9
{n1}
{n1, n2}
{n2}
{n1, n2}
{n1, n2}
{n2, n3, n4}
{n4}
{n3, n4}
{n3, n4}
ƒ Procedure of initializing a tree
» Start from leaf nodes (level-0)
» At each level l, find non-overlapping groups of similar
nodes with frequent pattern mining
142
Updating Similarities Between Nodes
ƒ The initial similarities can seldom capture the relationships
between objects
ƒ Iteratively update similarities
» Similarity between two nodes is the average similarity between
objects linked with them
1
4
5
0
ST2
2
3
6
7
8
sim(na,nb) =
average similarity between
9
c
ST1
d
e
f
l m n
o p
g
q r
t
and
13
14
takes O(3x2) time
h
s
11
12
10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
z
10
k
u v w
x
y
143
Aggregation-Based Similarity Computation
0.2
4
0.9
5
1.0 0.8
10
11
0.9
1.0
13
14
12
a
b
ST2
ST1
For each node nk ∈ {n10, n11, n12} and nl ∈ {n13, n14}, their pathbased similarity simp(nk, nl) = s(nk, n4)·s(n4, n5)·s(n5, nl).
∑
sim(n , n ) =
12
a
b
k =10
s(nk , n4 )
3
∑
⋅ s(n , n ) ⋅
14
4
5
l =13
s(nl , n5 )
2
= 0.171
takes O(3+2) time
After aggregation, we reduce quadratic time computation to linear
time computation.
144
Computing Similarity with Aggregation
Average similarity
and total weight
sim(na, nb) can be computed
from aggregated similarities
a:(0.9,3)
0.2
4
10
11
12
b:(0.95,2)
5
13
a
14
b
sim(na, nb) = avg_sim(na,n4) x s(n4, n5) x avg_sim(nb,n5)
= 0.9 x 0.2 x 0.95 = 0.171
To compute sim(na,nb):
ƒ Find all pairs of sibling nodes ni and nj, so that na linked with ni and nb
with nj.
ƒ Calculate similarity (and weight) between na and nb w.r.t. ni and nj.
ƒ Calculate weighted average similarity between na and nb w.r.t. all such
pairs.
145
Adjusting SimTree Structures
n1
n4
0.8
n7
0.9
n2
n5
n7 n8
n3
n6
n9
ƒ After similarity changes, the tree structure also
needs to be changed
» If a node is more similar to its parent’s sibling, then move
it to be a child of that sibling
» Try to move each node to its parent’s sibling that it is
most similar to, under the constraint that each parent
node can have at most c children
146
Complexity
For two types of objects, N in each, and M linkages between them.
Time
Space
Updating similarities
O(M(logN)2)
O(M+N)
Adjusting tree
structures
O(N)
O(N)
LinkClus
O(M(logN)2)
O(M+N)
SimRank
O(M2)
O(N2)
147
Experiment: Email Dataset
ƒ
ƒ
ƒ
ƒ
„
F. Nielsen. Email dataset.
www.imm.dtu.dk/~rem/data/Email-1431.zip
370 emails on conferences, 272 on jobs,
and 789 spam emails
Accuracy: measured by manually labeled
data
Accuracy of clustering: % of pairs of objects
in the same cluster that share common label
Approach
Accuracy time (s)
LinkClus
0.8026
1579.6
SimRank
0.7965
39160
ReCom
0.5711
74.6
F-SimRank
0.3688
479.7
CLARANS
0.4768
8.55
Approaches compared:
„
SimRank (Jeh & Widom, KDD 2002): Computing pair-wise similarities
„
SimRank with FingerPrints (F-SimRank): Fogaras & R´acz, WWW 2005
„
„
pre-computes a large sample of random paths from each object and uses
samples of two objects to estimate SimRank similarity
ReCom (Wang et al. SIGIR 2003)
„
Iteratively clustering objects using cluster labels of linked objects
148
Clustering – Sub-Topics
ƒ
What is Cluster Analysis?
ƒ
Types of Data in Cluster Analysis
ƒ
A Categorization of Major Clustering Methods
ƒ
Partitioning Methods
ƒ
Hierarchical Methods
ƒ
Density-Based Methods
ƒ
Grid-Based Methods
ƒ
Model-Based Methods
ƒ
Clustering High-Dimensional Data
ƒ
Constraint-Based Clustering
ƒ
Link-based clustering
ƒ
Outlier Analysis
149
What Is Outlier Discovery?
ƒ What are outliers?
» The set of objects are considerably dissimilar from the
remainder of the data
» Example: Sports: Michael Jordon, Wayne Gretzky, ...
ƒ Problem: Define and find outliers in large data
sets
ƒ Applications:
» Credit card fraud detection
» Telecom fraud detection
» Customer segmentation
» Medical analysis
150
Outlier Discovery:
Statistical Approaches
Assume a model underlying distribution that generates
data set (e.g. normal distribution)
ƒ Use discordancy tests depending on
» data distribution
» distribution parameter (e.g., mean, variance)
» number of expected outliers
ƒ Drawbacks
» most tests are for single attribute
» In many cases, data distribution may not be known
151
Outlier Discovery: Distance-Based Approach
ƒ Introduced to counter the main limitations
imposed by statistical methods
» We need multi-dimensional analysis without knowing
data distribution
ƒ Distance-based outlier: A DB(p, D)-outlier is an
object O in a dataset T such that at least a
fraction p of the objects in T lies at a distance
greater than D from O
ƒ Algorithms for mining distance-based outliers
[Knorr & Ng, VLDB’98]
» Index-based algorithm
» Nested-loop algorithm
» Cell-based algorithm
152
Density-Based Local Outlier Detection
ƒ M. M. Breunig, H.-P. Kriegel, R. Ng, J.
Sander. LOF: Identifying Density-Based
Local Outliers. SIGMOD 2000.
ƒ Distance-based outlier detection is based
on global distance distribution
„
ƒ It encounters difficulties to identify outliers
if data is not uniformly distributed
„
ƒ Ex. C1 contains 400 loosely distributed
points, C2 has 100 tightly condensed
Need the concept of local
outlier
Local outlier factor (LOF)
„ Assume outlier is not
crisp
„ Each point has a LOF
points, 2 outlier points o1, o2
ƒ Distance-based method cannot identify o2
as an outlier
153
Outlier Discovery: Deviation-Based Approach
ƒ Identifies outliers by examining the main
characteristics of objects in a group
ƒ Objects that “deviate” from this description are
considered outliers
ƒ Sequential exception technique
» simulates the way in which humans can distinguish
unusual objects from among a series of supposedly
like objects
ƒ OLAP data cube technique
» uses data cubes to identify regions of anomalies in
large multidimensional data
154
Agenda
11
Session
Session Overview
Overview
22
Clustering
Clustering
33
Summary
Summary and
and Conclusion
Conclusion
155
Summary
ƒ Cluster analysis groups objects based on their similarity
and has wide applications
ƒ Measure of similarity can be computed for various types
of data
ƒ Clustering algorithms can be categorized into partitioning
methods, hierarchical methods, density-based methods,
grid-based methods, and model-based methods
ƒ Outlier detection and analysis are very useful for fraud
detection, etc. and can be performed by statistical,
distance-based or deviation-based approaches
ƒ There are still lots of research issues on cluster analysis
156
Problems and Challenges
ƒ Considerable progress has been made in scalable
clustering methods
» Partitioning: k-means, k-medoids, CLARANS
» Hierarchical: BIRCH, ROCK, CHAMELEON
» Density-based: DBSCAN, OPTICS, DenClue
» Grid-based: STING, WaveCluster, CLIQUE
» Model-based: EM, Cobweb, SOM
» Frequent pattern-based: pCluster
» Constraint-based: COD, constrained-clustering
ƒ Current clustering techniques do not address all the
requirements adequately, still an active area of research
157
References (1)
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
R. Agrawal, J. Gehrke, D. Gunopulos, and P. Raghavan. Automatic subspace
clustering of high dimensional data for data mining applications. SIGMOD'98
M. R. Anderberg. Cluster Analysis for Applications. Academic Press, 1973.
M. Ankerst, M. Breunig, H.-P. Kriegel, and J. Sander. Optics: Ordering points
to identify the clustering structure, SIGMOD’99.
Beil F., Ester M., Xu X.: "Frequent Term-Based Text Clustering", KDD'02
M. M. Breunig, H.-P. Kriegel, R. Ng, J. Sander. LOF: Identifying DensityBased Local Outliers. SIGMOD 2000.
M. Ester, H.-P. Kriegel, J. Sander, and X. Xu. A density-based algorithm for
discovering clusters in large spatial databases. KDD'96.
M. Ester, H.-P. Kriegel, and X. Xu. Knowledge discovery in large spatial
databases: Focusing techniques for efficient class identification. SSD'95.
D. Fisher. Knowledge acquisition via incremental conceptual clustering.
Machine Learning, 2:139-172, 1987.
D. Gibson, J. Kleinberg, and P. Raghavan. Clustering categorical data: An
approach based on dynamic systems. VLDB’98.
V. Ganti, J. Gehrke, R. Ramakrishan. CACTUS Clustering Categorical Data
Using Summaries. KDD'99.
158
References (2)
ƒ D. Gibson, J. Kleinberg, and P. Raghavan. Clustering categorical data:
An approach based on dynamic systems. In Proc. VLDB’98.
ƒ S. Guha, R. Rastogi, and K. Shim. Cure: An efficient clustering
algorithm for large databases. SIGMOD'98.
ƒ S. Guha, R. Rastogi, and K. Shim. ROCK: A robust clustering
algorithm for categorical attributes. In ICDE'99, pp. 512-521, Sydney,
Australia, March 1999.
ƒ A. Hinneburg, D.l A. Keim: An Efficient Approach to Clustering in
Large Multimedia Databases with Noise. KDD’98.
ƒ A. K. Jain and R. C. Dubes. Algorithms for Clustering Data. Printice
Hall, 1988.
ƒ G. Karypis, E.-H. Han, and V. Kumar. CHAMELEON: A Hierarchical
Clustering Algorithm Using Dynamic Modeling. COMPUTER, 32(8):
68-75, 1999.
ƒ L. Kaufman and P. J. Rousseeuw. Finding Groups in Data: an
Introduction to Cluster Analysis. John Wiley & Sons, 1990.
ƒ E. Knorr and R. Ng. Algorithms for mining distance-based outliers in
large datasets. VLDB’98.
159
References (3)
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
ƒ
G. J. McLachlan and K.E. Bkasford. Mixture Models: Inference and
Applications to Clustering. John Wiley and Sons, 1988.
P. Michaud. Clustering Techniques. Future Generation Computer Systems,
13, 1997.
R. Ng and J. Han. Efficient and effective clustering method for spatial data
mining. VLDB'94.
L. Parsons, E. Haque and H. Liu, Subspace Clustering for High Dimensional
Data: A Review, SIGKDD Explorations, 6(1), June 2004
E. Schikuta. Grid clustering: An efficient hierarchical clustering method for
very large data sets. Proc. 1996 Int. Conf. on Pattern Recognition,.
G. Sheikholeslami, S. Chatterjee, and A. Zhang. WaveCluster: A multiresolution clustering approach for very large spatial databases. VLDB’98.
A. K. H. Tung, J. Han, L. V. S. Lakshmanan, and R. T. Ng. Constraint-Based
Clustering in Large Databases, ICDT'01.
A. K. H. Tung, J. Hou, and J. Han. Spatial Clustering in the Presence of
Obstacles, ICDE'01
H. Wang, W. Wang, J. Yang, and P.S. Yu. Clustering by pattern similarity in
large data sets, SIGMOD’ 02.
W. Wang, Yang, R. Muntz, STING: A Statistical Information grid Approach to
Spatial Data Mining, VLDB’97.
T. Zhang, R. Ramakrishnan, and M. Livny. BIRCH : An efficient data
clustering method for very large databases. SIGMOD'96.
160
Clustering with User-Specified Constraints
ƒ Tung, Han, Lakshmanan, and Ng. Constraint-Based
Clustering in Large Databases, ICDT'01
ƒ Example: Locating k delivery centers, each serving at least
m valued customers and n ordinary ones
ƒ Proposed approach
» Find an initial “solution” by partitioning the data set into k groups
and satisfying user-constraints
» Iteratively refine the solution by micro-clustering relocation (e.g.,
moving δ micro-clusters from cluster Ci to Cj) and “deadlock”
handling (breaking the micro-clusters when necessary)
» Efficiency is improved by micro-clustering
161
Assignments & Readings
ƒ Readings
» Chapter 7
ƒ Individual Project #2
» Due 04/15/10
ƒ Group Project Proposal
» Due 04/08/10
162
Next Session: Data Mining Applications (Part I)
163