Visualization Blaz Zupan Faculty of Computer & Info Science University of Ljubljana, Slovenia

Visualization

Blaz Zupan

Faculty of Computer & Info ScienceUniversity of Ljubljana, Slovenia

Visual Data Mining

• Basic idea– visual presentation of the data– gain insight & generate hypothesis– draw conclusions– directly interact with data

• Include human in the data exploration process– use her/his flexibility– creativity– general knowledge

Benefits of Visualization

• involvement of the user• results are intuitive

– no need for understanding complex mathematical or statistical algorithms or parameters

• provision of qualitative overview of the data– can isolate specific patterns for further

quantitative analysis

• can deal with non-homogenous, noisy data

Visual Exploration Paradigm

Overview first, zoom & filter, and then details on demand.


Visual Exploration Paradigm



Classification

one-dimensional

two-dimensional

multi-dimensional

text, web content

networks

other (e.g. algorithms/software, ...)

Data Type

Interaction & DistortionTechnique

VisualizationTechnique

Standard

Projection

Filtering

Zoom

Distortion

Link & Brush

Standard 2D/3D Display

Geometrically Transformed Display

Iconic Display

Dense Pixel Display

Stacked Display

from D Keim & M Ward: Visualization, in Intelligent Data Analysis, M Berthold & DJ Hand (eds), Springer, 2003.

Data: One-Dimensional

R Bellazzi: Mining Biomedical Time Series by Combining Structural Analysis and Temporal Abstractions,In Proc. of AMIA 1998.

Data: Two-Dimensional

MineSet’s Map Visualizer.

Data: Multi-Dimensional

Data: Text • Galaxies visualization

• Uses the “night sky” visualization to represent a set of documents

• One document – one star

• Stars clustered together represent related documents

• Includes analytical tools to investigate groups and time-based trends, query contents

From Inspire (TM) Software, seewww.pnl.gov/infoviz/technologies.html

Data: Text • ThemeView (TM)• Topics or themes

of text documents shown in relief map of a natural terrain

• The height of a peek relates to the strength of the topic


Data: Text • Theme River (TM)

• Identification of time related trends and patterns

• Themes represented as colored streams

• The width of the stream relates to the collective strength of a theme


Data: Networks

E. coli metabolic network (colors denote predominant biochemical class of metabolites)Ravasz et al., Science 297, 30 Aug 2002.

S. cerevisiae gene interaction networkTong et al., Science 303, 6 Feb 2004.

V Batagelj, A Mrvar: Pajek @ vlado.fmf.uni-lj.si/pub/networks/pajek/

Data: Tree Hierarchies

Unix home directory Selected detail

Kleiberg et al.: Botanic Visualization of Huge Hierarchies, In InfoVis, 2001.

Classification

one-dimensional

two-dimensional

multi-dimensional

text, web content

networks


Data Type



Standard

Projection

Filtering

Zoom

Distortion

Link & Brush



Iconic Display

Dense Pixel Display

Stacked Display

Standard 2D/3D

• x-y (x-y-z) plots

• bar charts• line graphs• histograms• maps

Standard 2D/3D



Standard 2D/3D



Geom.-Transformed Displays

• includes several classes of visualizations

• projection pursuit, finding “interesting transformations” of multi-dim data set

• scatterplot matrix

• parallel coordinates

Iconic Displays

W Horn et al.: Metaphor graphics to visualize ICU data over time, In IDAMAP 1998.

Dense Pixel Displays

DA Keim et al.: Recursive Pattern: A technique for visualizing very large amounts of dataProc. Visualization 95, pages 279-286, 1995.

Dense Pixel Displays

Ankerst et al.: Circle Segments: A technique for visually exploring large multidimensional data sets.In Proc. Visualization 96, Hot Topic Session, 1996.

Stacked Displays

J LeBlanc et al.: Exploring n-dimensional databases. In Proc. Visualization 90, pages 230-239, 1990.

• an example is dimensional stacking

• embed one coordinate system within the other

• e.g. two attributes in one system, then another two when drilling down

Stacked Displays

Decision table visualization from SGI’s MineSet

Stacked Displays

Mosaic display in Orange.

Classification

one-dimensional

two-dimensional

multi-dimensional

text, web content

networks


Data Type



Standard

Dynamic Projection

Filtering

Zoom

Distortion

Link & Brush



Iconic Display

Dense Pixel Display

Stacked Display

Interaction Techniques

• Dynamic projection– dynamically change the

projections to explore multi-dimensional data sets

– projection pursuit, which finds well-separated clusters in scatterplot

• Interactive Filtering– browsing, can be

difficult for big data sets– querying, need to

specify a subset

• Zooming• Distortion

– e.g., fisheye view

• Brushing and linking– requires well-

integrated system for visualization

– selection from one visualization is fed into another one, selected instances highlighted in some way

Distortion

GW Furnas: Generalized Fisheye Views, Human Factors in Computing Systems CHI ‘86 Conference Proceedings, 16-23. 1986.

Distortion

From M Grobelnik, P Krese, D Mladenic: Project Intelligence (http://pi.ijs.si)

Distortion


Distortion


Brushing & Linking

Integration ofVisualization & Data Mining

1. Visualization techniques can be applied before (or independently) of DM

2. DM can be used to find patterns (or data subsets) that are further visualized

3. DM is interactive, users use visualization to guide the pattern search

4. Visualization of data mining models

Regression Tree

Regression tree visualization in SGI’s MineSet.

Classification Tree

Classification tree visualization in Orange.

Brushing: Trees & Scatter Plots

Sieve Diagram (Classification)

Nomograms

Intelligent Data Visualization

• Use an established visualization technique, but search for– interesting subset of attributes– interesting subset of data instances– interesting projection (how to use selected

attributes in visualization)

• All these to find “interesting” visualization

• Removes the burden for the user to find such visualizations by hand

Arrangement for Circle Segments

M Ankerst: Visual data mining with pixel-oriented techniques, In Proc. KDD, 2001.

VizRank

Conclusion

• Clarity of presentation• Aesthetics• Navigation & Interaction

• In data with many dimensions, tools are needed to find only “interesting” visualizations

Documents

Visualization Blaz Zupan Faculty of Computer & Info Science University of Ljubljana, Slovenia