Outlier-Preserving Focus Context Visualization in Parallel Coordinates

1
Outlier-Preserving FocusContext Visualization in
Parallel Coordinates
Matej Novotný Comenius University Bratislava, Slovakia Helwig Hauser VRVis Research Center Vienna, Austria
2
Our goal

• A parallel coordinates visualization that
• Employs FocusContext
• Handles outliers
• Renders effectively

3
Overview

• Motivation
• Abstraction, FocusContext
• Outliers
• Workflow
• Binning
• Context
• Benefits
• Bonus!
• Results and conclusions

4
Parallel Coordinates

• Insight into multidimensional data
• Correlations, Groups, Outliers

5
Parallel Coordinates

• Insight into multidimensional data
• Correlations, Groups, Outliers

6
Parallel Coordinates

• Insight into multidimensional data
• Correlations, Groups, Outliers

7
Large data visualization

• Large data cause clutter in visualization
• 16.000 records

8
Large data visualization

• Transparency used to decrease clutter
• 16.000 records

9
Large data visualization

• Transparency used to decrease clutter ?
• 32.000 records

10
Large data visualization

• Transparency used to decrease clutter ??
• 64.000 records

11
Large data visualization

• Transparency used to decrease clutter ???
• 100.000 records

12
Large data visualization

• Transparency used to decrease clutter ???
• Do these records belong to the main trend?

13
Data abstraction

• Density-based representation of data
• Trends are clearly visible

16 bins
14
Related work

• Hierarchical Parallel Coordinates(Fua et al.,
  1999)
• Visual representationof clusters
• Smooth transparency
• Cluster centersemphasized

15
Related work

• Revealing Structure within Clustered Parallel
  Coordinates Displays (Johansson et al., 2005)
• Textures, density
• Transfer functions
• Clusters
• Outliers

16
Outliers

• Different, sparse, rare
• Why should we care?
• Investigation (special cases in simulations)
• Security (network intrusion, suspicious
  activity)
• Detect errors in data acquisition
• Outliers can be considered in
• Data space
• Screen space

17
Outliers
Outliers are like kids. If you leave them
unattendedthey either get lostor they break
stuff.
18
Outliers

• Avoid losing them in visualization
• e.g. due to transparency or abstraction
• Improve data abstraction or FC
• e.g. remove outliers from clustering

19
Workflow

20
Workflow

21
Step 1 Binning

• 2D binning
• Density-based rep.
• Screen-oriented
• Low memory demandscompared to nD
• Every segmentseparately
• Result bin map

22
Benefits of binning?

• Operations no longer dependon the size of the
  input
• Information is preserved
• Variable precision of binning
• Variable precision of visual output
• Fine binning does not destroy details
• Larger bins can be producedfrom finer bins

128x128 bins
23
Step 2 Outlier detection

• Various criteria can be employed
• e.g. isolated bins, median filter

64x64 bin map
32x32 bin mapmedian filter
32x32 bin mapisolated bins
24
Step 3 Generating Context

• Outliers ? opaque lines
• Binned trends ? quads
• Population mapped to color intensity
• No blending
• Low visual complexity
• Rendering order according to population

8 bins
25
Step 4 Add Focus

8 bins
26
Benefits

• Operations performed on bin maps
• Reduced complexity
• Results coherent with visual output
• More operations feasible e.g. Clustering
• Outliers handled separately
• Increased information value
• Clearer context
• Output-sensitive implementation
• View divided into layers and segments

27
Results

• Large data can be rendered and explored
• 3 millions records, 16 dimensions, 32 bins
• Binned in 30 sec, rendered instantly (3Ghz,64bit)

28

• BONUS!
• Clustering

29
Clustering, step 0

• Apply Gaussian to smooth out the bin map
• Segmentation data, Green vs Darkness

30
Clustering, further steps

• Start with the highest population
• Decrease the population threshold
• Old clusters grow
• New clusters emerge

50
20
10
0
31
Clustering results

R B G D S H
32
Clustering results

R B G D S H
33
Clustering results

R B G D S H
34
Clustering results

R B G D S H
35
Conclusions

• Data abstraction based on density rep.
• Data operations - outlier detection, clustering
• FocusContext
• Variable context precision
• Outliers preserved
• Much clearer view for large data
• Screen-oriented and output-sensitive
• Interactive visualization of large data

36
Acknowledgements

• K-Plus
• Vega grant 1/3083/06.
• AVL List GmbH - data
• Juergen Platzer
• Prof. Peter Filzmoser
• Harald Piringer
• Michael Wohlfahrt

37
Thank you for your attention!