Agglomerative - anuradhasrinivas

Download Report

Transcript Agglomerative - anuradhasrinivas

K-Means Algorithm
 Each cluster is represented by the mean value of the
objects in the cluster
 Input
: set of objects (n), no of clusters (k)
 Output
: set of k clusters
 Algo
 Randomly select k samples & mark them a initial cluster
 Repeat


Assign/ reassign in sample to any given cluster to which it is
most similar depending upon the mean of the cluster
Update the cluster’s mean until No Change.
K-Means (graph)
 Step1:
 Step2:
Form k centroids, randomly
Calculate distance between centroids and
each object
 Use Euclidean’s law do determine min distance:
d(A,B) = (x2-x1)2 + (y2-y1)2
 Step3:
 Step4:
C=
Assign objects based on min distance to k
clusters
Calculate centroid of each cluster using
(x1+x2+…xn , y1+y2+…yn)
n
n
 Go to step 2.
 Repeat until no change in centroids.
K-Mediod (PAM)




Also called Partitioning Around Mediods.
Step1:
choose k mediods
Step2:
assign all points to closest mediod
Step3:
form distance matrix for each cluster and
choose the next best mediod. i.e., the point closest to
all other points in cluster
 go to step2.
 Repeat until no change in any mediods
What are Hierarchical
Methods?
 Groups data objects into a tree of clusters
 Classified as
 Agglomerative (Bottom-up)
 Divisive (Top-Bottom)
 Once a merge or split decision is made it cannot be
backtracked
Types of hierarchical clustering
 Agglomerative (Bottom-up) AGNES
 Places each object into a cluster and merges atomic
clusters into larger clusters
 They differ in the definition of intercluster similarity
 Divisive: (Top-Bottom) DIANA
 All objects are initially in one cluster
 Subdivides the cluster into smaller and smaller pieces,
until each object forms a cluster of its own or satisfies
some termination condition
 In both of the above methods the termination
condition is the number of clusters
Dendogram
Level 4
Level 3
Level 2
Level 1
Level 0
Measures of Distance
 Minimum distance – Nearest Neighbor- single linkage
–minimum spanning tree
 Maximum distance – Farthest neighbor clustering
algorithm – complete linkage
 Mean distance - avoids outlier sensitivity problem
 Average distance : can handle categorical as well as
numeric data
Euclidean Distance
Agglomerative Algorithm
 Step1:
 Step2:
Make each object as a cluster
Calculate the Euclidean distance from every
point to every other point. i.e., construct a
Distance Matrix
Identify two clusters with shortest distance.
 Step3:
 Merge them
 Go to Step 2
 Repeat until all objects are in one cluster
Agglomerative Algorithm
Approaches
 Single Link:
 Quite simple
 Not very efficient
 Suffers from chain effect
 Complete Link
 More compact than those found using the single link
technique
 Average Link
Simple Example
Item
E
A
C
B
D
E
0
1
2
2
3
A
1
0
2
5
3
C
2
2
0
1
6
B
2
5
1
0
3
D
3
3
6
3
0
Another Example
 Find single link technique to find clusters in the given
database.
1
2
3
4
5
6
X
Y
0.4
0.53
0.22
0.38
0.35
0.32
0.26
0.19
0.08
0.41
0.45
0.3
Plot given data
Identify two nearest clusters
Repeat process until all objects
in same cluster
Average link
 Average distance matrix
Construct a distance matrix
1
2
3
4
5
1
0
2
0.24
0
3
0.22
0.15
0
4
0.37
0.2
0.15
0
5
0.34
0.14
0.28
0.29
0
6
0.23
0.25
0.11
0.22
0.39
6
0
Divisive Clustering
 All items are initially placed in one cluster
 The clusters are repeatedly split in two until all items
are in their own cluster
1
A
B
2
C
E
3
1
D
Difficulties in Hierarchical
Clustering
 Difficulties regarding the selection of merge or split
points
 This decision is critical because the further merge or
split decisions are based on the newly formed clusters
 Method does not scale well
 So hierarchical methods are integrated with other
clustering techniques to form multiple-phase
clustering
Types of hierarchical clustering
techniques
 BIRCH-Balanced Iterative Reducing and Clustering
using hierarchies
 ROCK: Robust clustering with links, explores the
concept of links
 CHAMELEON: hierarchical clustering algorithm using
dynamic modeling
Outlier Analysis
 Outliers are data objects, which are different from or
inconsistent with the remaining set of data
 Outliers can be caused because of
 Measurement or execution error
 Result of inherent data variability
 Can be used in fraud detection
 Outlier detection and analysis is referred to as outlier
mining.
Applications of outlier mining
 Fraud detection
 Customized marketing for identifying the spending
behavior of customers with extremely low or high
incomes.
 Medical analysis for finding unusual responses to
various medical treatments.
What is outlier mining?
 Given a set of n data points or objects and k, the
expected number of outliers find the top k objects that
are dissimilar, exceptional or inconsistent with respect
to remaining data
 There are two subproblems
 Define what data can be considered as inconsistent in a
given data set
 Method to mine the outliers
Methods of outlier detection
 Statistical approach
 distance-based approach
 Density-based local outlier approach
 Deviation-based approach
Statistical Distribution
 Identifies outliers with respect to a discordancy test
 Discordancy test examines a working hypothesis and
an alternative hypothesis
 It verifies whether an object oi, is significantly large in
relation to the distribution F.
 This helps in accepting the working hypothesis or
rejecting it (alternative distribution)
 Inherent alternative distribution
 Mixture alternative distribution
 Slippage alternative distribution
Procedures for detecting
outliers
 Block procedures: All suspect objects are treated as
outliers or all of then are accepted as consistent
 Consecutive procedures: object that is least likely to be
an outlier is tested first. If it is found to be an outlier
then all of the more extreme values are also considered
as outliers. Else the next most extreme object is tested
and so on
Questions in Clustering