Вы находитесь на странице: 1из 17

# Unsupervised Learning Basic Concepts

Outline

## Part 1: Basic Concepts of data clustering

 Non-Supervised Learning and Clustering
: Problem formulation – cluster analysis
: Taxonomies of Clustering Techniques
: Data types and Proximity Measures
: Difficulties and open problems

## Part 2: Clustering Algorithms

 Hierarchical methods
: Clustering Based on Dissimilarity Increments Criteria

## Pattern Recognition – Decision Making

 Supervised Learning
: training samples, labeled by their category membership, are used
to design a classifier

## . Labeled training patterns

. Labels represent true categories of patterns

 Unsupervised Learning
: Based on a collection of samples without being told their
categories

. Learn the number of classes and the structure of each class using
similarity between unlabeled training patterns

. Datamining

## 2 From Single Clustering to Ensemble Methods - April 2009

1
Unsupervised Learning Basic Concepts

## Unsupervised Learning / Clustering

 Unsupervised Learning
: Learn the structure of multidimensional patterns
. Mixture Densities
– Gaussian Mixture Decomposition
» The probability structure is known with the exception of the values of
the parameters

 Clustering Procedures
: Find subclasses
. Data description in terms of clusters or groups of data points that
possess strong internal similarities

 Typical applications:
. As a stand-alone tool to get insight into data distribution
. As a preprocessing step for other algorithms

## Unsupervised Learning Basic Concepts

Cluster Analysis
Organize data into sensible groupings (either as a grouping of patterns or a
hierarchy of groups)

 Clustering
: The process of grouping a set of objects into classes of similar objects
(extracting hidden structure from data)
 Cluster
: A collection of objects that are similar to one another within the same
cluster and are dissimilar to the objects in other clusters

## 4 From Single Clustering to Ensemble Methods - April 2009

2
Unsupervised Learning Basic Concepts

Shape Clustering

## Cistern from MR brain images

The main cluster is drawn using multicolor dots, secondary clusters are drawn in red,
green and magenta.

Duta, Jain and Jolly, “Automatic Construction of 2-D Shape Models”, IEEE PAMI, May 2001

Shape Clustering

## Cistern from MR brain images

The main cluster is drawn using multicolor dots, secondary clusters are drawn in red,
green and magenta.

Duta, Jain and Jolly, “Automatic Construction of 2-D Shape Models”, IEEE PAMI, May 2001

## 6 From Single Clustering to Ensemble Methods - April 2009

3
Unsupervised Learning Basic Concepts

## Identification of Writing Styles

 122,000 “online” characters written by 100 writers
 Lexemes are identified by clustering data within each character class into subclasses:
a string matching measure used to calculate distance between 2 characters

Connell and Jain, “Writer Adaptation for Online Handwriting Recognition”, IEEE PAMI, Mar 2002

## Segmentation of Natural Scenes

Hermes, Zoller, Bumannn, “Parametric Distributional Clustering for Image Segmentation”, ECCV 2002

## 8 From Single Clustering to Ensemble Methods - April 2009

4
Unsupervised Learning Basic Concepts

What is a Cluster?

 A set of entities which are alike; entities from different clusters are
not alike
 An aggregation of points such that the distance between any two
points in a cluster is less than the distance between any point in the
cluster and any point not in it.
 A relatively high density of points, surrounded by a relatively low
density of points

## Two main strategies:

Hierarchical Methods
:Propose a sequence of nested data
partitions in a hierarchical structure

Partitional Methods
:Organize patterns into a small
number of clusters
. K-means
. Spectral clustering

## 10 From Single Clustering to Ensemble Methods - April 2009

5
Unsupervised Learning Basic Concepts

## Taxonomy of Clustering Approaches

Clustering Principles:

 Compactness
: K-means
: Histogram clustering
: Pairwise data clustering

 Connectedness
: Dissimilarity Increments
: Mean Shift clustering

 Separation
: Normalized Cut
: Spectral clustering

## Taxonomy of Clustering Approaches

Clustering Principles:

 Compactness
: K-means
: Histogram clustering
: Pairwise data clustering

 Connectedness
: Dissimilarity Increments
: Mean Shift clustering

 Separation
: Normalized Cut
: Spectral clustering

## 12 From Single Clustering to Ensemble Methods - April 2009

6
Unsupervised Learning Basic Concepts

## Taxonomy of Clustering Approaches

Clustering Principles:

 Compactness
: K-means
: Histogram clustering
: Pairwise data clustering

 Connectedness 10

9

: Dissimilarity Increments
8

## : Mean Shift clustering

7

5
 Separation 4
: Normalized Cut
3
: Spectral clustering
2

1
2 3 4 5 6 7 8 9 10

## Taxonomy of Clustering Approaches

Approaches:

 Model-based
: Patterns can be given a simple and compact description in terms of
. Parametrical distribution -- Parametric density approaches (Mixture
models)
. A representative element, such as a centroid, median (central
clustering, square-error clustering, k-means, k-medoids) – or
multiple prototypes per cluster (CURE) -- Prototype-based methods
. Some geometrical primitives (lines, planes, circles, curves,
surfaces) – Shape fitting approaches
: These approaches assume particular cluster shapes, partitions
being in general obtained as a result of an optimization process
using a global criterion

## 14 From Single Clustering to Ensemble Methods - April 2009

7
Unsupervised Learning Basic Concepts

## Taxonomy of Clustering Approaches

 Graph-theoretical
: Mostly explored in hierarchical methods that can be represented
graphically as a tree or dendrogram
. Divisive approaches (ex. Based on Minimum Spanning Tree)
: View clustering as a graph partitioning problem

##  Non parametric density-based

: Attempt to identify high density clusters separated by low density
regions (local cluster criterion, such as density-connected points)
(valley seeking clustering algorithms)
. DBSCAN, OPTICS, DENCLUE, CLIQUE
. Discover clusters of arbitrary shape

## Data Types in Clustering Problems

Data representations:
Vector data: n vectors in Rd

proximity matrix

## :All types of data may be

proximity measures

## 16 From Single Clustering to Ensemble Methods - April 2009

8
Unsupervised Learning Basic Concepts

##  Distances are normally used to measure the similarity or

dissimilarity between two data objects

 Metrics:
: Positivity: d(a, b) >0 and d(a, b)=0 , a=b
: Symmetry property: d(a,b)=d(b,a).
: Triangle inequality: d(a,c)· d(a,b) + d(b,c).

## Metric Models in Feature Spaces

 Minskowski distance:

## 18 From Single Clustering to Ensemble Methods - April 2009

9
Unsupervised Learning Basic Concepts

## Metric Models in Feature Spaces

 Absolute Value Metric, Manhattan Distance or City-block (r = 1)
d
d M (a, b)  d1 (a, b)   bi  ai
i 1

## Constant Manhattan distance curves:

. Reduced computational time; does not penalize much the features with
higher dissimilarity
. In R2: dist1((x1,y1),(x2,y2))=|x2-x1|+|y2-y1| , city-block:
It is not possible to make short-cuts through corners: it counts the number of
blocks that is necessary to pass in order to move from one corner to
another

## Metric Models in Feature Spaces

 Euclidean Distance:
d
d e ( a, b)  d 2 ( a, b)   b  a 
2
i i
i 1

. R2: dist2((x1,y1),(x2,y2))=((x2-x1)2+(y2-y1)2)1/2.

## . Emphasizes more features with higher dissimilarity.

 Mahalanobis Distance

d Mahalanobis ( x, y)   x  y   1  x  y 
T

## 20 From Single Clustering to Ensemble Methods - April 2009

10
Unsupervised Learning Basic Concepts

## Dissimilarity for Sequences of Symbols

 Dissimilarity based on String Editing operations

.
.
.
.

 The Levensthein distance between two strings s1, s2 2 *, DL(s1, s2), is
defined as the minimum number of editing operations needed in order to
transform s1 into s2.

## Dissimilarity for Sequences of Symbols

 The Weighted Levensthein distance between two strings s1, s2 2 *, is
defined by

where

## 22 From Single Clustering to Ensemble Methods - April 2009

11
Unsupervised Learning Basic Concepts

## String matching. In (b), diagonal path elements represent substitutions,

vertical segments correspond to insertions, and horizontal
segments correspond to deletions.

##  Normalized Edit Distance

Marzal and Vidal, “Computation of normalized edit distance and applications”, IEEE PAMI, 1993

## 24 From Single Clustering to Ensemble Methods - April 2009

12
Unsupervised Learning Basic Concepts

##  Dissimilarity based on Error-Correcting Parsing [Fu]

: distance between strings based on the modelling of string structure
by means of grammars and on the concept of error-correcting
parsing
: the distance between a string and a reference string is given by the
error-correcting parser as the weighted Levensthein distance
between the string and the nearest (in terms of edit operations)
string generated by the grammar inferred from the reference string
(thus exhibiting a similar structure):

ECP distance

## 26 From Single Clustering to Ensemble Methods - April 2009

13
Unsupervised Learning Basic Concepts

##  Dissimilarity based on Error-Correcting Parsing [Fu]

: distance between strings based on the modelling of string structure
by means of grammars and on the concept of error-correcting
parsing
: the distance between a string and a reference string is given by the
error-correcting parser as the weighted Levensthein distance
between the string and the nearest (in terms of edit operations)
string generated by the grammar inferred from the reference string
(thus exhibiting a similar structure):

##  Grammar Complexity-based Similarity

The basic idea is that, if two sentences are structurally similar, then
their joint description will be more compact than their isolated
description due to sharing of rules of symbol composition; the
compactness of the representation is quantified by the grammar
complexity, and the similarity is measured by the ratio of decrease
in grammar complexity

## where C(Gsi) denotes grammar complexity.

Fred, “Clustering of Sequences using a Minimum Grammar Complexity Criterion”, ICGI 1996
Fred. “Similarity measures and clustering of string patterns”. In Dechang Chen and Xiuzhen Cheng, editors,
Pattern Recognition and String Matching, Kluwer Academic, 2002,

## 28 From Single Clustering to Ensemble Methods - April 2009

14
Unsupervised Learning Basic Concepts

RDGC Similarity

##  Grammar Complexity-based Similarity– RDGC

Let G=(VN ,  , R, ) be a context-free grammar, where VN,  are the sets
of nonterminal and terminal symbols, respectively,  is the grammar’s
start symbol and R is the set of productions written in the form:

## Let  2 (VN  )*, be a grammatical sentence of length n, in which the

symbols a1, a2, … , am appear k1, k2, … , km times, respectively. The
complexity of the sentence, C(), is given by [Fu]

## 30 From Single Clustering to Ensemble Methods - April 2009

15
Unsupervised Learning Basic Concepts

##  Minimum Code Length-based Similarity

: Based on Solomonoff’s code: a string is represented by a triplet

## where a coded string is obtained in an iterative procedure where, in

each step, intermediate codes are produced by defining sequences of
two symbols, which are represented by special symbols, and rewriting
the sequences using them. Compact codes are produced when
sequences exhibit local or distant inter-symbol interactions.
. Code length: sum of the lengths of the descriptions of the three part
code above
: Extension to sets of strings

Fred and Leitão, “A Minimum Code Length Technique for Clustering of Syntactic Patterns”, ICPR 1996
Fred. “Similarity measures and clustering of string patterns”. In Dechang Chen and Xiuzhen Cheng, editors,
Pattern Recognition and String Matching, Kluwer Academic, 2002,

##  Minimum Code Length-based Similarity

: The basic idea is that global compact codes are produced by
considering the inter-symbol dependencies on the ensemble of the
strings. The quantification of this reduction in code length forms the
basis of the similarity measure designated by Normalized Ratio of
decrease in code length - NRDCL

with

## 32 From Single Clustering to Ensemble Methods - April 2009

16
Unsupervised Learning Basic Concepts

##  Discovery of clusters with arbitrary shape

 Ability to deal with different types of attributes
 Scalability
 Minimal requirements for domain knowledge to determine input
parameters
 Insensitivity to the order of input records
 Ability to deal with noisy data
 High dimensionality

## Unsupervised Learning Basic Concepts

Issues in Clustering

##  Which similarity measure and features to use?

 How many clusters?
 Which is the “best” clustering method?
 Are the individual clusters and the partition valid?
 How to choose algorithmic parameters?

## K-means clustering of uniform data K-means using Euclidean (blue) and

(k=4) Mahalanobis distance (k=2) (red)

17