Segmentação ( Clustering ) (baseado nos slides do Han)
-
Upload
maris-koch -
Category
Documents
-
view
32 -
download
4
description
Transcript of Segmentação ( Clustering ) (baseado nos slides do Han)
Segmentação (Segmentação (ClusteringClustering))(baseado nos slides do (baseado nos slides do
Han)Han)
SAD Tagus 2004/05 H. Galhardas
Non-supervised Non-supervised Learning: Learning: Cluster Cluster
AnalysisAnalysis What is Cluster Analysis?What is Cluster Analysis? Types of Data in Cluster AnalysisTypes of Data in Cluster Analysis A Categorization of Major Clustering MethodsA Categorization of Major Clustering Methods Partitioning MethodsPartitioning Methods Hierarchical MethodsHierarchical Methods Summary Summary
SAD Tagus 2004/05 H. Galhardas
What is Cluster What is Cluster Analysis?Analysis?
ClusterCluster: a collection of data objects: a collection of data objects Similar to one another within the same cluster Dissimilar to the objects in other clusters
Cluster analysisCluster analysis Grouping a set of data objects into clusters
Clustering is Clustering is unsupervised classificationunsupervised classification: no : no predefined classespredefined classes
Typical applicationsTypical applications As a stand-alone tool to get insight into data distribution As a preprocessing step for other algorithms
SAD Tagus 2004/05 H. Galhardas
General Applications of General Applications of Clustering Clustering
Pattern RecognitionPattern Recognition Spatial Data Analysis Spatial Data Analysis
create thematic maps in GIS by clustering feature spaces detect spatial clusters and explain them in spatial data
mining Image ProcessingImage Processing Economic Science (especially market research)Economic Science (especially market research) WWWWWW
Document classification Cluster Weblog data to discover groups of similar access
patterns
SAD Tagus 2004/05 H. Galhardas
Examples of Clustering Examples of Clustering ApplicationsApplications
MarketingMarketing: Help marketers discover distinct groups in their : Help marketers discover distinct groups in their customer bases, and then use this knowledge to develop customer bases, and then use this knowledge to develop targeted marketing programstargeted marketing programs
Land useLand use: Identification of areas of similar land use in an : Identification of areas of similar land use in an earth observation databaseearth observation database
InsuranceInsurance: Identifying groups of motor insurance policy : Identifying groups of motor insurance policy holders with a high average claim costholders with a high average claim cost
City-planningCity-planning: Identifying groups of houses according to : Identifying groups of houses according to their house type, value, and geographical locationtheir house type, value, and geographical location
Earth-quake studies:Earth-quake studies: Observed earth quake epicenters Observed earth quake epicenters should be clustered along continent faultsshould be clustered along continent faults
SAD Tagus 2004/05 H. Galhardas
What Is Good Clustering?What Is Good Clustering?
A A good clusteringgood clustering method will produce high quality method will produce high quality
clusters withclusters with
high intra-class similarity
low inter-class similarity
The The qualityquality of a clustering result depends on both the of a clustering result depends on both the
similarity measure used by the method and its similarity measure used by the method and its
implementation.implementation.
The The qualityquality of a clustering method is also measured by its of a clustering method is also measured by its
ability to discover some or all of the ability to discover some or all of the hiddenhidden patterns patterns..
SAD Tagus 2004/05 H. Galhardas
Requirements of Requirements of Clustering in Data Mining Clustering in Data Mining
ScalabilityScalability Ability to deal with different types of attributesAbility to deal with different types of attributes Discovery of clusters with arbitrary shapeDiscovery of clusters with arbitrary shape Minimal requirements for domain knowledge to determine Minimal requirements for domain knowledge to determine
input parametersinput parameters Able to deal with noise and outliersAble to deal with noise and outliers Insensitive to order of input recordsInsensitive to order of input records High dimensionalityHigh dimensionality Incorporation of user-specified constraintsIncorporation of user-specified constraints Interpretability and usabilityInterpretability and usability
SAD Tagus 2004/05 H. Galhardas
Data Data StructuresStructures
Data matrixData matrix (two modes)
Dissimilarity matrixDissimilarity matrix (one mode)
npx...
nfx...
n1x
...............ip
x...if
x...i1
x
...............1p
x...1f
x...11
x
0...)2,()1,(
:::
)2,3()
...ndnd
0dd(3,1
0d(2,1)
0
SAD Tagus 2004/05 H. Galhardas
Measure the Quality Measure the Quality of Clusteringof Clustering
Dissimilarity/Similarity metricDissimilarity/Similarity metric: Similarity is expressed in terms : Similarity is expressed in terms of a distance function, which is typically a metric:of a distance function, which is typically a metric:dd((i, ji, j))
There is a separate There is a separate “quality” function“quality” function that measures the that measures the “goodness” of a cluster.“goodness” of a cluster.
The definitions of distance functions are usually very The definitions of distance functions are usually very different for interval-scaled, boolean, categorical, ordinal and different for interval-scaled, boolean, categorical, ordinal and ratio variables.ratio variables.
WeightsWeights should be associated with different variables based should be associated with different variables based on applications and data semantics.on applications and data semantics.
It is hard to define “similar enough” or “good enough” It is hard to define “similar enough” or “good enough” the answer is typically highly subjective.
SAD Tagus 2004/05 H. Galhardas
Type of data in clustering Type of data in clustering analysisanalysis
Interval-scaled variables:Interval-scaled variables:
Binary variables:Binary variables:
Nominal, ordinal, and ratio variables:Nominal, ordinal, and ratio variables:
Variables of mixed types:Variables of mixed types:
SAD Tagus 2004/05 H. Galhardas
Interval-valued variablesInterval-valued variables
Standardize dataStandardize data
Calculate the mean absolute deviation:
where
Calculate the standardized measurement (z-score)
Using mean absolute deviation is more robust than using Using mean absolute deviation is more robust than using
standard deviation standard deviation
.)...21
1nffff
xx(xn m
|)|...|||(|121 fnffffff
mxmxmxns
f
fifif s
mx z
SAD Tagus 2004/05 H. Galhardas
Similarity and Dissimilarity Similarity and Dissimilarity Between ObjectsBetween Objects
DistancesDistances are normally used to measure the are normally used to measure the similaritysimilarity
or or dissimilaritydissimilarity between two data objects between two data objects
Some popular ones include: Some popular ones include: Minkowski distanceMinkowski distance
where where ii = ( = (xxi1i1, , xxi2i2, …, , …, xxipip) and) and j j = ( = (xxj1j1, , xxj2j2, …, , …, xxjpjp) are two ) are two
pp-dimensional data objects, and -dimensional data objects, and qq is a positive integer is a positive integer
If If qq = = 11, , dd is is Manhattan distanceManhattan distance
pp
jx
ix
jx
ix
jx
ixjid )||...|||(|),(
2211
||...||||),(2211 pp jxixjxixjxixjid
SAD Tagus 2004/05 H. Galhardas
Similarity and Dissimilarity Similarity and Dissimilarity Between Objects (Cont.)Between Objects (Cont.)
If qIf q = = 22,, d d is is Euclidean distanceEuclidean distance::
PropertiesProperties::d(i,j) 0d(i,i) = 0d(i,j) = d(j,i)d(i,j) d(i,k) + d(k,j)
Also, one can use weighted distance, Also, one can use weighted distance, parametric Pearson product moment parametric Pearson product moment correlation, or other disimilarity measurescorrelation, or other disimilarity measures
)||...|||(|),( 22
22
2
11 pp jx
ix
jx
ix
jx
ixjid
SAD Tagus 2004/05 H. Galhardas
Partitioning Algorithms: Partitioning Algorithms: Basic ConceptBasic Concept
Partitioning methodPartitioning method: Construct a partition of a : Construct a partition of a database database DD of of nn objects into a set of objects into a set of kk clusters clusters
Given a Given a kk, find a partition of , find a partition of k clusters k clusters that that optimizes the chosen partitioning criterionoptimizes 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
SAD Tagus 2004/05 H. Galhardas
The The K-MeansK-Means Clustering Clustering MethodMethod
Given Given kk, the , the k-meansk-means algorithm is algorithm is
implemented in four steps:implemented in four steps:
1. Partition objects into k nonempty subsets
2. 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)
3. Assign each object to the cluster with the nearest
seed point
4. Go back to Step 2, stop when no more new
assignment
SAD Tagus 2004/05 H. Galhardas
The The K-MeansK-Means Clustering Clustering MethodMethod
ExampleExample
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
K=2
Arbitrarily choose K object as initial cluster center
Assign each objects to most similar center
Update the cluster means
Update the cluster means
reassignreassign
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
SAD Tagus 2004/05 H. Galhardas
Comments on the Comments on the K-MeansK-Means MethodMethod
StrengthStrength Relatively efficientRelatively efficient: : OO((tkntkn), where ), where nn is # objects, is # objects, kk is # is #
clusters, and clusters, and t t is # iterations. Normally, is # iterations. Normally, kk, , tt << << nn.. Comparing: PAM: O(k(n-k)2 ), CLARA: O(ks2 + k(n-k))
CommentComment Often terminates at a Often terminates at a local optimumlocal optimum. The . The global global
optimumoptimum may be found using techniques such as: may be found using techniques such as: deterministic deterministic
annealingannealing and and genetic algorithmsgenetic algorithms
WeaknessWeakness 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
SAD Tagus 2004/05 H. Galhardas
Variations of the Variations of the K-MeansK-Means MethodMethod
A few variants of the A few variants of the k-meansk-means which differ in which differ in
Selection of the initial k means
Dissimilarity calculations
Strategies to calculate cluster means
Handling categorical data: Handling categorical data: k-modesk-modes (Huang’98)(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
SAD Tagus 2004/05 H. Galhardas
What is the problem of k-What is the problem of k-Means Method?Means Method?
The k-means algorithm is sensitive to outliers !The k-means algorithm is sensitive to outliers ! Since an object with an extremely large value may substantially
distort the distribution of the data.
K-MedoidsK-Medoids: Instead of taking the : Instead of taking the meanmean value of the object value of the object
in a cluster as a reference point, in a cluster as a reference point, medoidsmedoids can be used, can be used,
which is the which is the most centrally locatedmost centrally located object in a cluster. object in a cluster.
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 100
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
SAD Tagus 2004/05 H. Galhardas
TheThe K K--MedoidsMedoids Clustering Clustering MethodMethod
Find Find representativerepresentative objects, called objects, called medoids, in , in
clustersclusters
PAMPAM (Partitioning Around Medoids, 1987) (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
SAD Tagus 2004/05 H. Galhardas
Typical k-medoids algorithm Typical k-medoids algorithm (PAM)(PAM)
Total Cost = 20
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
K=2
Arbitrary choose k object as initial medoids
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
Assign each remaining object to nearest medoids Randomly select a
nonmedoid object,Oramdom
Compute total cost of swapping
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
Total Cost = 26
Swapping O and Oramdom
If quality is improved.
Do loop
Until no change
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
SAD Tagus 2004/05 H. Galhardas
PAM (Partitioning Around PAM (Partitioning Around Medoids) (1987)Medoids) (1987)
Use real object to represent the clusterUse real object to represent the cluster
1. Select k representative objects arbitrarily
2. For each pair of non-selected object h and selected
object i, calculate the total swapping cost TCih
3. 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
4. Repeat steps 2-3 until there is no change
SAD Tagus 2004/05 H. Galhardas
Cost function for k-medoidsCost function for k-medoids
SAD Tagus 2004/05 H. Galhardas
PAM ClusteringPAM Clustering
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
j
ih
t
Cjih = 0
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
t
i h
j
Cjih = d(j, h) - d(j, i)
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
h
i t
j
Cjih = d(j, t) - d(j, i)
0
1
2
3
4
5
6
7
8
9
10
0 1 2 3 4 5 6 7 8 9 10
t
ih j
Cjih = d(j, h) - d(j, t)
SAD Tagus 2004/05 H. Galhardas
What is the problem What is the problem with PAM?with PAM?
Pam is Pam is more robustmore robust than k-means in the than k-means in the
presence of noise and outliers because a medoid presence of noise and outliers because a medoid
is less influenced by outliers or other extreme is less influenced by outliers or other extreme
values than a meanvalues than a mean
Pam works efficiently for small data sets but Pam works efficiently for small data sets but
does not does not scale wellscale well for large data sets. for large data sets. O(k(n-k)2 ) for each iteration, where n is # of data,k is #
of cluster
SAD Tagus 2004/05 H. Galhardas
SummarySummary
Cluster analysisCluster analysis groups objects based on their groups objects based on their similaritysimilarity and has wide applicationsand has wide applications
Measure of similarity can be computed for Measure of similarity can be computed for various types various types of dataof data
Clustering algorithms can be Clustering algorithms can be categorizedcategorized into partitioning into partitioning methods, hierarchical methods, density-based methods, methods, hierarchical methods, density-based methods, grid-based methods, and model-based methodsgrid-based methods, and model-based methods
Outlier detectionOutlier detection and analysis are very useful for fraud and analysis are very useful for fraud detection, etc. and can be performed by statistical, detection, etc. and can be performed by statistical, distance-based or deviation-based approachesdistance-based or deviation-based approaches
There are still lots of research issues on cluster analysis, There are still lots of research issues on cluster analysis, such as such as constraint-based clusteringconstraint-based clustering
SAD Tagus 2004/05 H. Galhardas
Other Classification Other Classification MethodsMethods
k-nearest neighbor classifierk-nearest neighbor classifier case-based reasoningcase-based reasoning Genetic algorithmGenetic algorithm Rough set approachRough set approach Fuzzy set approachesFuzzy set approaches
SAD Tagus 2004/05 H. Galhardas
Instance-Based Instance-Based MethodsMethods
Instance-based learningInstance-based learning: : Store training examples and delay the processing
(“lazy evaluation”) until a new instance must be classified
Typical approachesTypical approaches k-nearest neighbor approach
Instances represented as points in a Euclidean space.
Locally weighted regressionConstructs local approximation
Case-based reasoningUses symbolic representations and knowledge-
based inference
SAD Tagus 2004/05 H. Galhardas
The The kk-Nearest Neighbor -Nearest Neighbor AlgorithmAlgorithm
All instances correspond to All instances correspond to pointspoints in the in the n-dimensional space.n-dimensional space.
The nearest neighbor are defined in The nearest neighbor are defined in terms of terms of Euclidean distanceEuclidean distance..
The target function could be discrete- or The target function could be discrete- or real- valued.real- valued.
SAD Tagus 2004/05 H. Galhardas
The The kk-Nearest Neighbor -Nearest Neighbor AlgorithmAlgorithm
For discrete-valued, the For discrete-valued, the kk-NN returns the -NN returns the most common valuemost common value among the k among the k training examples nearest totraining examples nearest to xxqq. .
Vonoroi diagramVonoroi diagram: the decision surface : the decision surface induced by 1-NN for a typical set of induced by 1-NN for a typical set of training examples.training examples.
.
_+
_ xq
+
_ _+
_
_
+
.
..
. .
SAD Tagus 2004/05 H. Galhardas
Discussion (1)Discussion (1)
The k-NN algorithm for The k-NN algorithm for continuous-valuedcontinuous-valued target target functionsfunctions Calculate the mean values of the k nearest neighbors
Distance-weighted nearest neighbor algorithmDistance-weighted nearest neighbor algorithm Weight the contribution of each of the k neighbors
according to their distance to the query point xq,giving greater weight to closer neighbors
Similarly, for real-valued target functionsw
d xq xi 1
2( , )
SAD Tagus 2004/05 H. Galhardas
Discussion (2)Discussion (2)
Robust to Robust to noisy datanoisy data by averaging k-nearest by averaging k-nearest neighborsneighbors
Curse of dimensionalityCurse of dimensionality: distance between : distance between neighbors could be dominated by irrelevant neighbors could be dominated by irrelevant attributes. attributes. To overcome it, axes stretch or elimination of the least
relevant attributes.
SAD Tagus 2004/05 H. Galhardas
BibliografiaBibliografia
Data Mining: Concepts and TechniquesData Mining: Concepts and Techniques, J. , J. Han & M. Kamber, Morgan Kaufmann, Han & M. Kamber, Morgan Kaufmann, 2001 (Sect. 7.7.1 e Cap. 8)2001 (Sect. 7.7.1 e Cap. 8)