2. visualization in data mining

nice

1. Mobile no - [phone]
1

2. 2
Motivation
Visualization for Data Mining
• Huge amounts of information
• Limited display capacity of output devices
Visual Data Mining (VDM) is a new approach for
exploring very large data sets, combining traditional
mining methods and information visualization techniques.

3. 3
Levels of VDM
No or very limited integration
Corresponds to the application of either traditional information
visualization or automated data mining methods.
Loose integration
Visualization and automated mining methods are applied
sequentially.
The result of one step can be used as input for another step.
Full integration
Automated mining and visualization methods applied in parallel.
Combination of the results.

4. 4
Methods of Data Visualization
Different methods are available for visualization of data
based on type of data
Data can be
Univariate
Bivariate
Multivariate

5. 5
Univariate data
Measurement of single quantitative variable
Characterize distribution
Represented using following methods
Histogram
Pie Chart

6. 6
Histogram

7. 7
Pie Chart

8. 8
Bivariate Data
Constitutes of paired samples of two quantitative
variables
Variables are related
Represented using following methods
Scatter plots
Line graphs

9. 9
Scatter plots

10. 10
Line graphs

11. 11
Multivariate Data
Multi dimensional representation of multivariate
data
Represented using following methods
Icon based methods
Pixel based methods
Dynamic parallel coordinate system

12. 12
Icon based Methods

13. 13
Pixel Based Methods
 Approach:
 Each attribute value is represented by one colored pixel
(the value ranges of the attributes are mapped to a
fixed color map).
 The values of each attribute are presented in separate
sub windows.
Examples:
 Dense Pixel Displays

14. 14
Dense Pixel Display
Approach:
 Each attribute value is represented by one colored
pixel (the value ranges of the attributes are mapped
to a fixed color map).
 Different attributes are presented in separate sub
windows.

15. 15
Visual Data Mining: Framework and
Algorithm Development
Ganesh, M., Han, E.H., Kumar, V., Shekar, S., &
Srivastava, J. (1996).
Working Paper. Twin Cities, MN: University of Minnesota,
Twin Cities Campus.

16. 16
References
http://coblitz.codeen.org:3125/citeseer.ist.psu.edu/cache/papers/cs/10335/ftp:
zSzzSzftp.cs.umn.eduzSzdeptzSzuserszSzkumarzSzdatavis.pdf/ganesh96vis
ual.pdf
http://www.ailab.si/blaz/predavanja/ozp/gradivo/2002-Keim-Visualization%20in
%20DM-IEEE%20Trans%20Vis.pdf
http://www.geocities.com/anand_palm/

17. 17
Abstract
VDM refers to refers to the use of visualization techniques in Data
Mining process to
Evaluate
Monitor
Guide
This paper provides a framework for VDM via the loose coupling of
databases and visualization systems.
The paper applies VDM towards designing new algorithms that can
learn decision trees by manually refining some of the decisions made
by well known algorithms such as C4.5.

18. 18
Components of VQLBCI
The three major components of VQLBCI are
Visual Representations, Computations and
Events.

19. 19
Visual Development of Algorithms
Most interesting use of visual data mining is the
development of new insights and algorithms.
The figure below shows the ER diagram for
learning classification decision trees.
This model allows the user to monitor the quality
and impact of decisions made by the learning
procedure.
Learning procedure can be refined interactively
via a visual interface.

20. 20
ER diagram for the search space of decision tree
learning algorithm

21. 21
General Framework
Learning a classification decision tree from a training data
set can be regarded as a process of searching for the best
decision tree that meets user-provided goal constraints.
The problem space of this search process consists of
Model Candidates, Model Candidate Generator and Model
Constraints.
Many existing classification-learning algorithms like C4.5
and CDP fit nicely within this search framework. New
learning algorithms that fit user’s requirements can be
developed by defining the components of the problem
space.

22. 22
General Framework
Model Candidate corresponds to the partial
classification decision tree. Each node of the
decision tree is a Model Atom
Search process is the process of finding a final
model candidate such that it meets user goal
specifications.
Model Candidate Generator transforms the
current model candidate into a new model
candidate by selecting one model atom to expand
from the expandable leaf model atoms.
Model Constraints (used by Model Candidate
Generator) provide controls and boundaries to the
search space.

23. 23
Search Process

24. 24
Acceptability Constraint
Model Constraints consist of Acceptability constraints,
Expandability constraints and a Data-Entropy calculation
function.
Acceptability constraint predicate specifies when a model
candidate is acceptable and thus allows search process to
stop. EX:
A1) Total no of expandable leaf model atoms = 0.
A2) Overall error rate of the model candidate <= acceptable error
rate.
A3) Total number of model atoms in the model candidate>=
maximal allowable tree size.
A1 is used in C4.5 and CDP

25. 25
Expandability Constraint
An Expandability constraint predicate specifies
whether a leaf model atom is expandable or not.
EX:
C4.5 uses E1 and E2
CDP uses E2 and E3

26. 26
Traversal Strategy
Traversal strategy ranks expandable leaf model
atoms based on the model atom attributes. EX:
Increasing order of depth
Decreasing order of depth
Orders based on other model atom attributes.

27. 27
Steps in Visual Algorithm Development
No single algorithm is the best all the time,
performance is highly data dependent.
By changing different predicates of model
constraints, users can construct new
classification-learning algorithm.
This enables users to find an algorithm that works
the best on a given data set.
Two algorithms are developed : BF based on
Best First search idea and CDP+ which is a
modification of CDP

28. 28
BF
This algorithm is based on the Best-First search
idea.
For Acceptability criteria, it includes A1 and A2
with a user specified acceptable error rate.
The Traversal strategy chosen is T3
In Best-First, expandable leaf model atoms are
ranked according to the decreasing order of the
number of misclassified training cases. (local error
rate * size of subset training data set)
The traversal strategy will expand a model atom
that has the most misclassified training cases,
thus reducing the overall error rate the most.

29. 29
CDP +
CDP+ is a modification of CDP
CDP has dynamic pruning using expandability
constraint E3.
Here, the depth is modified according to the size
of the training data set of the model atom.
We set
B is the branching factor of the decision tree, t is
the size of training data set belonging to model
atom, T is the whole training data set.

30. 30
Comparison of different classification learning
algorithms

31. 31
Experiment
The new BF and CDP+ algorithms are compared
with the C4.5 and CDP algorithms.
Various metrics are selected to compare the
efficiency, accuracy and size of final decision
trees of the classification algorithm.
The generation efficiency of the nodes is
measured in terms of the total number of nodes
generated.
To compare accuracy of the various algorithms,
the mean classification error on the test data sets
have been computed.

32. 32
Classification error for 10 data sets

33. 33
Nodes generated for 10 data
sets

34. 34
Final decision tree size

35. 35
Results/Conclusion
CDP has accuracy comparable to C4.5 while
generating considerably fewer nodes.
CDP+ has accuracy comparable to C4.5 while
generating considerably fewer nodes.
CDP+ outperformed CDP in error rate and
number of nodes generated.
Considering all performance metrics together,
CDP+ is the best overall algorithm.
Considering classification accuracy alone, C4.5P
is the winner.

36. 36
Conclusion
Different datasets require different algorithms for
best results.
Diverse user requirements put different
constraints on the final decision tree.
The experiment shows that Interactive Visual
Data Mining Framework can help find the most
suitable algorithm for a given data set and group
of user requirements.

37. 37
Data Mining for Selective Visualization
of Large Spatial Datasets
Proceedings of 14th IEEE International Conference on Tools with
Artificial Intelligence
(ICTAI'02),  2002.
Washington (November 2002), DC, USA,
Shashi Shekhar, Chang-Tien Lu, Pusheng Zhang, Rulin Liu
Computer Science & Engineering Department
University of Minnesota

38. 38
References
http://citeseer.ist.psu.edu/cache/papers/cs/27216/http:zSzzSzwww-
users.cs.umn.eduzSzzCz7EctluzSzPaperTalkFilezSzits02.pdf/shekha
r02cubeview.pdf
http://www.cs.umn.edu/Research/shashi-group/
http://www.cs.umn.edu/Research/shashi-group/Book/sdb-chap1.pdf
http://www.cs.umn.edu/research/shashi-group/alan_planb.pdf
http://coblitz.codeen.org:3125/citeseer.ist.psu.edu/cache/papers/cs/27
637/http:zSzzSzwww-
users.cs.umn.eduzSzzCz7EpushengzSzpubzSzkdd2001zSzkdd.pdf/s
hekhar01detecting.pdf

39. 39
Basic Terminology
Spatial databases
Alphanumeric data + geographical cordinates
Spatial mining
Mining of spatial databases
Spatial datawarehouse
Contains geographical data
Spatial outliers
Observations that appear to be inconsistent with the
remainder of that set of data

40. 40
Spatial Cluster

41. 41
Contribution
Propose and implement the CubeView
visualization system
General data cube operations
Built on the concept of spatial data warehouse to
support data mining and data visualization
Efficient and scalable spatial outlier detection
algorithms

42. 42
Challenges in spatial data mining
Classical data mining - numbers and categories.
Spatial data –
more complex and
extended objects such as points, lines and polygons.
Second, classical data mining works with explicit
inputs, whereas spatial predicates and attributes
are often implicit.
 Third, classical data mining treats each input
independently of other inputs.

43. 43
Application Domain
The Traffic Management Center - Minnesota
Department of Transportation (MNDOT) has a
database to archive sensor network.
Sensor network includes
about nine hundred stations
each of which contains one to four loop detector
Measurement of Volume and occupancy.
Volume is # vehicles passing through station in 5-
minute interval
Occupancy is percentage of time station is occupied
with vehicles

44. 44
Basic Concepts
Spatial Data Warehouse
Spatial Data Mining
Spatial Outliers Detection

45. 45
Spatial Data Warehouse
Employs data cube structure
Outputs - albums of maps.
Traffic data warehouse
Measures - volume and occupancy
Dimensions - time and space.

46. 46
Spatial Data Mining
Process of discovering interesting and useful but
implicit spatial patterns.
key goal is to partially ‘automate’ knowledge
discovery
Search for “nuggets” of information embedded in
very large quantities of spatial data.

47. 47
Spatial Outliers Detection
Suspiciously deviating observations
Local instability
Each Station
Spatial attributes – time, space
Non spatial attributes – volume, occupancy

48. 48
Basic Structure – CubeView

49. 49
CubeView Visualization System
Each node in cube – a visualization style
S - Traffic volume of station at all times.
T
TD
– Time of the day
T
DW
– Day of the week
ST
TD
– Daily traffic volume of each station
T
TD
T
DW
S– Traffic volume at each station at different times
on different days

50. 50
Dimension Lattice

51. 51
CubeView Visualization System

52. 52
CubeView Visualization System

53. 53
CubeView Visualization System

54. 54
Data Mining Algorithms for
Visualization
Problem Definition
Given a spatial graph G ={ S , E }
S - s1, s2, s3, s4……..
E – edges (neighborhood of stations)

f ( x ) - attribute value for a data record
N ( x )- fixed cardinality set of neighbors of x
 ) - Average attribute value of x neighbors
S( x ) - difference of the attribute value of each data
object and the average attribute value of neighbors.

55. 55
Data Mining Algorithms for
Visualization
Problem Definition cont…
S( x ) - difference of the attribute value of each data
object and the average attribute value of neighbors.
Test for detecting an outlier
 confidence level threshold θ



56. 56
Data Mining Algorithms for
Visualization
Few points
First, the neighborhood can be selected based on a fixed
cardinality or a fixed graph distance or a fixed Euclidean distance.
Second, the choice of neighborhood aggregate function can be
mean, variance, or auto-correlation.
 Third, the choice for comparing a location with its neighbors can
be either just a number or a vector of attribute values.
Finally, the statistic for the base distribution can be selected as
normal distribution.

57. 57
Data Mining Algorithms for
Visualization
Algorithms
Test Parameters Computation(TPC) Algorithm
Route Outlier Detection(ROD) Algorithm

58. 58
Data Mining Algorithms for
Visualization

59. 59
Data Mining Algorithms for
Visualization

60. 60
Data Mining Algorithms for
Visualization

61. 61
Software
http://www.cs.umn.edu/research/shashi-
group/vis/traffic_volumemap2.htm
http://www.cs.umn.edu/research/shashi-
group/vis/DataCube.htm

62. 62
Visualization and Data Mining
techniques
Thank you!!!!