The document discusses the complexities and methodologies involved in visualizing high-dimensional data, highlighting the challenges of accurate representation and potential for misinterpretation. Key concepts include various visualization techniques such as pair plots, t-SNE, and principal component analysis, as well as the importance of understanding human perception in data interpretation. The presentation emphasizes that while data visualization is inherently flawed, it can still be incredibly useful when done thoughtfully.

1. Visualizing
High-dimensional
Data
Dr. Stefan Kühn
Lead Data Scientist
code.talks - 30.09.2016

2. Overview
• Data Visualization as you might know it
• Main Properties of Graphics (and Humans)
• A short story about Charts
• Pair Plots and Correlations
• Data Visualization as you might not know it
• Fundamental Problems
• SVD, t-SNE and other approximations
• Principal Components and Principal Curves
• Parallel Coordinates
• The Grand Tour
2

3. Takeaways - hopefully ;-)
• Data Visualization is complicated
• There is always an approximation
• There is always a bias
• There is always a misinterpretation
• Data Visualization is simple
• Lots of packages available
• Lots of studies + literature
• Lots of examples to learn from
3

4. Data Visualization Basics
4

5. The Modes of Perception
5
X
X
X
X
X
X
X
O
O
O O
O
O
O
O
X
X
X
X
O
O
O
O
X
X
X
X
X
X
X
X
O
O
O O
O
O
O
O
X
X X
X
X
O
O
O
Fast Slow
Find the outlier

6. The Modes of Perception
6
• Pre-attentive
• fast
• parallel processing
• effortless
• Pattern recognition
• semi-fast
• governed by laws of per by
• Attentive
• slow
• sequential
• high effort (attention a very limited resource)

7. Main Properties of Graphics
7
Category Example
Position
Shape
Size
Color
Orientation (Line)
Length (Line)
Type and Size (Line)
Brightness

8. Main Properties of Graphics Humans
8
Category Amount of pre-attentive information
Position very high
Shape ———
Size approx. 4
Color approx. 8
Orientation (Line) approx. 4
Length (Line) ———
Type and Size (Line) ———
Brightness approx. 8

9. Pre-attentive perception
9
• Position
• fast
• effective
• high number of different positions
• Color
• use with care
• Shape
• Orientation
Pre-attentive perception is effortless.
Exploit this as much as you can.

10. Pattern detection
10
„It is interesting to note that our brain […]
subconsciously always prefers meaningful
situations and objects.“
• Emergence
• Reiification
• Multi-stability
• Invariance
Pattern detection can be trained.
Exploit this for frequent visualizations.

11. What is this?
11

12. What is this?
12

13. What is this?
13

14. Laws of Gestalt
14
„It is interesting to note that our brain, in
accordance with the laws of Gestalt,
subconsciously always prefers meaningful
situations and objects.“

15. Accuracy of Graphics
15
Square Pie vs Stacked Bar vs Pie vs Donut
What do you think?
https://eagereyes.org/blog/2016/a-reanalysis-of-a-study-about-square-pie-charts-from-2009

16. Demo
16

17. Beyond the Basics
17

18. Fundamental Problems
• No accurate method in higher dimensions
• Approximations methods
• „Simulated“ dimensions (color, size, shape)
• Animations?
• No notion of quality or accuracy for
Visualizations
• Information Theory?
• „Stability“?
All Visualizations are wrong, but some are useful.
18

19. Approximation methods
• Pair Plots
• Axis-aligned projections
• Interpretable in terms of original variables
• Singular Value Decomposition
• Optimal with respect to 2-norm (Euclidean norm)
and supremum norm
• Comes with an error estimate
• Other methods
• Stochastic Neighborhood Embedding (t-SNE)
• „Manifold Learning“
19

20. Demo
20

21. Manifold Learning Methods
• Locally Linear Embedding
• Neighborhood-preserving embedding
• Isomap
• quasi-isometric
• Multi-dimensional scaling
• quasi-isometric
• Spectral Embedding
• Spectral clustering based on similarity
• Stochastic Neighborhood Embedding (SNE, t-SNE)
• preserves probabilities
• Local Tangent Space Alignement (LTSA)
21

22. Local Tangent Space Alignement
22

23. Principal Components and Curves
• Principal Component Analysis
• orthogonal decomposition based on SVD
• linear in all variables
• tries to preserve variance
• Principal Curves
• minimize the Sum of Squared Errors with respect
to all variables (as PCA, preserve variance)
• nonlinear
• smooth
23

24. Principal Components and Curves
24

25. Parallel Coordinates
• Parallel Coordinates
• especially useful for high-dimensional data
• depends on ordering and scaling
25

26. Demo
26

27. The Grand Tour
• Animated sequence of 2-D projections
• https://en.wikipedia.org/wiki/
Grand_Tour_(data_visualisation)
• Asimov (1985): The grand tour: a tool for viewing
multidimensional data.
• Underlying idea
• Randomly generate 2-D projections (random
walk)
• Over time generate a dense subset of all
possible 2-D projections
• Optional: Follow a given path / guided tour
27

28. The Grand Tour
28

29. The Grand Tour
29

30. Demo
30

31. 31
Thanks a lot!
www.codecentric.de
blog.codecentric.de
stefan.kuehn@codecentric.de
datascience@codecentric.de