HỆ HỖ TRỢ QUYẾT ĐỊNH

Bài 9 (b): Clustering

 What is clustering analysis

2
What is clustering analysis
Clustering
• Clustering is the process of finding meaningful groups in
data
• For example, the customers of a company can be grouped
based on purchase behavior.
• The task of clustering can be used in two different classes of
applications:
• To describe a given dataset and
• as a preprocessing step for other data science algorithms.

4
Clustering to describe the data
• The most common application of clustering is to explore the
data and find all the possible meaningful groups in the data
• Clustering a company’s customer records can yield a few
groups in such a way that customers within a group are
more like each other than customers belonging to a
different group
• Applications:
• Marketing: Finding the common groups of customers
• Document clustering: One common text mining task is to
automatically group documents
• Session grouping: In web analytics, clustering is helpful to
understand common groups of clickstream patterns

5
Clustering for preprocessing
• Clustering to reduce dimensionality
• Clustering for object reduction

6
 Types of clustering techniques
• The clustering process seeks to find groupings in data, in
such a way that data points within a cluster are more similar
to each other than to data points in the other clusters
• One common way of measuring similarity is the Euclidean
distance measurement in 𝑛 − 𝑑𝑖𝑚𝑒𝑛𝑠𝑖𝑜𝑛𝑎𝑙 space

7
Taxonomy based on data point’s membership
• Exclusive or strict partitioning clusters: Each data object
belongs to one exclusive cluster
• Overlapping clusters: The cluster groups are not exclusive,
and each data object may belong to more than one cluster.
• Hierarchical clusters: Each child cluster can be merged to
form a parent cluster.
• Fuzzy or probabilistic clusters: Each data point belongs to all
cluster groups with varying degrees of membership from 0
to 1.

8
 Taxonomy by algorithmic approach
• Prototype-based clustering: In the prototype-based clustering,
each cluster is represented by a central data object, also called a
prototype (centroid clustering or center-based clustering)
• Density clustering:
• Each dense area can be assigned a cluster and the low-density area can
be discarded as noise
• Hierarchical clustering:
• Hierarchical clustering is a process where a cluster hierarchy is created
based on the distance between data points.
• The output of a hierarchical clustering is a dendrogram: a tree diagram
that shows different clusters at any point of precision which is specified
by the user
• Model-based clustering:
• Model-based clustering gets its foundation from statistics and
probability distribution models; this technique is also called
distribution-based clustering.
• Mixture of Gaussians is one of the model-based clustering techniques

9
K-Means Clustering
k-Means
• k-Means clustering is a prototype-based clustering method where
the dataset is divided into 𝑘 − 𝑐𝑙𝑢𝑠𝑡𝑒𝑟𝑠.
• User specifies the number of clusters (𝑘) that need to be grouped
in the dataset.
• The objective of k-means clustering is to find a prototype data
point for each cluster; all the data points are then assigned to the
nearest prototype, which then forms a cluster.
• The prototype is called as the centroid, the center of the cluster.
• The center of the cluster can be the mean of all data objects in
the cluster, as in k-means, or the most represented data object, as
in k-medoid clustering.
• The cluster centroid or mean data object does not have to be a
real data point in the dataset and can be an imaginary data point
that represents the characteristics of all the data points within
the cluster.

11
k-Means
• The data objects inside a partition belong to the cluster.
• These partitions are also called Voronoi partitions, and each
prototype is a seed in a Voronoi partition.

12
Algorithm
• The logic of finding k-clusters within a given dataset is rather
simple and always converges to a solution
• However, the final result in most cases will be locally
optimal where the solution will not converge to the best
global solution
• Step 1: Initiate Centroids
• The first step in k-means algorithm is to initiate k random centroids
• Step 2: Assign Data Points
• all the data points are now assigned to the nearest centroid to form
a cluster.
• Step 3: Calculate New Centroids
• New centroids are means of each clusters

13
Algorithm
• Step 4: Repeat Assignment and Calculate New Centroids
• assigning data points to the nearest centroid is repeated until all
the data points are reassigned to new centroids
• Step 5: Termination
• Step 3—calculating new centroids, and step 4—assigning data
points to new centroids, are repeated until no further change in
assignment of data points happens.

14
Some key issues
• Initiation: The final clustering grouping depends on the
random initiator and the nature of the dataset.
• Empty clusters: One possibility in k-means clustering is the
formation of empty clusters in which no data objects are
associated.
• Outliers: Since SSE (sum of squared errors) is used as an
objective function, k-means clustering is susceptible to
outliers
• Post-processing: Since k-means clustering seeks to be locally
optimal, a few post-processing techniques can be
introduced to force a new solution that has less SSE

15
Evaluation of Clusters
• Evaluation of clustering can be as simple as computing total
SSE
• Good models will have low SSE within the cluster and low
overall SSE among all clusters.
• SSE can also be referred to as the average within-cluster
distance and can be calculated for each cluster and then
averaged for all the clusters.
• Davies-Bouldin index is a measure of uniqueness of the
clusters and takes into consideration both cohesiveness of
the cluster (distance between the data points and center of
the cluster) and separation between the clusters

16
k-Means in RapidMiner

17
DBSCAN Clustering
DBSCAN
• A cluster can also be defined as an area of high
concentration (or density) of data objects surrounded by
areas of low concentration (or density) of data objects.
• A density-clustering algorithm identifies clusters in the data
based on the measurement of the density distribution in n-
dimensional space
• Specifying the number of the cluster parameters (𝑘) is not
necessary for density-based algorithms
• Thus, density-based clustering can serve as an important
data exploration technique
• Density can be defined as the number of data points in a
unit n-dimensional space

19
Algorithm
• Step 1: Defining Epsilon and
MinPoints
• Calculation of a density for all
data points in a dataset, with a
given fixed radius 𝜀 (epsilon).
• To determine whether a
neighborhood is high-density
or low-density, a threshold of
data points (MinPoints) will
have to be defined, above
which the neighborhood is
considered high-density
• Both 𝜀 and 𝑀𝑖𝑛𝑃𝑜𝑖𝑛𝑡𝑠 are
user-defined parameters
20
 Algorithm
• Step 2: Classification of Data Points
Core points: All the data points inside the high-
density region of at least one data point are
considered a core point. A high-density region
is a space where there are at least MinPoints
data points within a radius of 𝜀 for any data
point.

Border points: Border points sit on the

Noise points: Any point that is neither a core point nor border point is called a
noise point. They form a low-density region around the high density region.

21
Algorithm
• Step 3: Clustering
• Groups of core points form distinct clusters. If two core points are
within 𝜀 of each other, then both core points are within the same
cluster
• All these clustered core points form a cluster, which is surrounded
by low-density noise points
• A few data points are left unlabeled or associated to a default noise
cluster

22
Special Cases: Varying Densities
• The dataset has four distinct regions numbered from 1-4.
Region 1 is the high-density area A, regions 2 and 4 are of
mediumdensity B, and between them is region 3, which is
extremely low-density C

• The density threshold parameters are tuned in such a way

23
DBSCAN in RapidMiner

24
Self-Organizing Maps
SOM
• A self-organizing map (SOM) is a powerful visual clustering
technique that evolved from a combination of neural
networks and prototype-based clustering
• A key distinction in this neural network is the absence of an
output target function to optimize or predict, hence, it is an
unsupervised learning algorithm
• SOM methodology is used to project data objects from data
space, mostly in 𝑛 dimensions, to grid space, usually
resulting in two dimensions

26
Algorithm
• Step 1: Topology Specification
• Two-dimensional rows and columns with either a rectangular lattice
or a hexagonal lattice are commonly used in SOMs
• The number of centroids is the product of the number of rows and
columns in the grid

27
Algorithm
• Step 2: Initialize Centroids
• A SOM starts the process by initializing the centroids. The initial
centroids are values of random data objects from the dataset. This
is similar to initializing centroids in k-means clustering.
• Step 3: Assignment of Data Objects
• After centroids are selected and placed on the grid in the
intersection of rows and columns, data objects are selected one by
one and assigned to the nearest centroid.
• Step 4: Centroid Update

28
Algorithm
• Step 5: Termination
• The entire algorithm is continued until no significant centroid
updates take place in each run or until the specified number of run
count is reached
• a SOM tends to converge to a solution in most cases but doesn’t
guarantee an optimal solution
• Step 6: Mapping a New Data Object
• any new data object can be quickly given a location on the grid
space, based on its proximity to the centroids.
• The characteristics of new data objects can be further understood
by studying the neighbors.

29
SOM in RapidMiner

30
SOM in RapidMiner

31
 Example
The dataset has 186 records, one for each country, and four attributes in percentage
of GDP:
• relative GDP invested
• relative GDP saved
• government revenue, and
• current account balance

32
33
34
35