Data exploration

• Discover patterns, stories in your data.
• Should be interactive, or at least flexible
• Should be comprehensive
  • include every relevant aspect
  • include only potentially relevant aspects
• Should be focused on its objective and its effective execution
• Can be unpolished

x1	y1	x2	y2	x3	y3	x4	y4	p
10	8.0	10	9.1	10	7.5	8	6.6	d
8	7.0	8	8.1	8	6.8	8	5.8	d
13	7.6	13	8.7	13	12.7	8	7.7	d
9	8.8	9	8.8	9	7.1	8	8.8	d
11	8.3	11	9.3	11	7.8	8	8.5	d
14	10.0	14	8.1	14	8.8	8	7.0	d
6	7.2	6	6.1	6	6.1	8	5.2	d
4	4.3	4	3.1	4	5.4	19	12.5	d
12	10.8	12	9.1	12	8.2	8	5.6	d
7	4.8	7	7.3	7	6.4	8	7.9	d
5	5.7	5	4.7	5	5.7	8	6.8	d
9	7.5	9	7.5	9	7.5	9	7.5	mean
11	4.1	11	4.1	11	4.1	11	4.1	var

x1	y1	x2	y2	x3	y3	x4	y4	p
10	8.0	10	9.1	10	7.5	8	6.6	d
8	7.0	8	8.1	8	6.8	8	5.8	d
13	7.6	13	8.7	13	12.7	8	7.7	d
9	8.8	9	8.8	9	7.1	8	8.8	d
11	8.3	11	9.3	11	7.8	8	8.5	d
14	10.0	14	8.1	14	8.8	8	7.0	d
6	7.2	6	6.1	6	6.1	8	5.2	d
4	4.3	4	3.1	4	5.4	19	12.5	d
12	10.8	12	9.1	12	8.2	8	5.6	d
7	4.8	7	7.3	7	6.4	8	7.9	d
5	5.7	5	4.7	5	5.7	8	6.8	d
9	7.5	9	7.5	9	7.5	9	7.5	mean
11	4.1	11	4.1	11	4.1	11	4.1	var

Why?

• What is the task (i.g., exploration, presentation)?
• Is there a message to communicate, and if so, what is it?

What?

• What data is shown?
• How is the data pre-processed, filtered, abstracted?

How?

• Which idiom is used?
• What attributes and which interactions are selected?
• How are selected attributes encoded?

Dataset types

• Tables
• Networks
• Fields
• Geometry
• Sets

Variable types

• Categorical:
  a factor
• Ordinal:
  a factor with a natural order
• Quantitative:
  an arithmetic value

Order types

• sequential (e.g., body height)
• diverging (e.g., temperature variation)
• cyclic (e.g. time of day)

• Sums
• Ratios
• Means
• Standard deviations, errors, variances
• Correlations
• Test statistics
• Dimensionality reduction (e.g., PCA)
• Grouping, clustering \(\Rightarrow\) centroids

• linear dimensionality reduction technique
• reduced coordinate system captures the largest variation in the data
• purpose
  • exploratory data analysis
  • visualization
  • data preprocessing

Legends

• A legend is a tool to help explain a graph.
• It translates between the channel (e.g., color, shape) and the respective encoded value.

Labels

• Labels can replace legends to some extend.
• Link between marks and explanatory text.
• Can also be associated with additonal channels (e.g., font size/color).
• Very useful to highlight and pop out items.

Munzner, Tamara. Visualization analysis and design. CRC press, 2014

1. Aligned position
2. Unaligned position
3. Length
4. Angle
5. Circular areas
6. Rectangular areas

Encoding attributes with a certain channel, are the differences between the individual items perceptible to humans as intended?

For some channels a very limited number of “bins” can be distinguished.

Binning the channel might improve the distinguisability of perception, but reduces the resolution.

The range of values for an attribute should match the number of bins which can be distinguished.

Channels can be separable or integrated.

Some interactions are obvious
\(\rightarrow\) two attributes are encoded by vertical and horizontal position, the planar proximity is no more available as a channel for a third attribute.
\(\rightarrow\) two attributes via size and shape on the same mark

Some interactions are less obvious
\(\rightarrow\) color hue interferes with size since it changes the perception of the color.

How quickly can a highlighted item catch your eye?

Popout is also referred to as preattentive processing or tunable detection

For channels with a favorable popout property, this works independent of the number of distractor objects.

Popout does not depend on the channel but also the particular encoding of the channel.

• Grouping can be achieved by (from least to most effective)
  • Identity marks (e.g., hue, shapes) are hardest to perceive
  • Link marks (i.e., lines, area of containment) add clutter but are easy to perceive
  • Spatial separation is the strongest cue, but hammpers comparability
• Appropriate technique strongly depends on particular task

• Identity and magnitude channel
• Can be used in 1D and in 2D (vertical and horizontal position)
• Closely related to length (distance between to positions)
• Little interference with other channel (not necessarily true the other way around)
• Most effective channel as magnitude and identity channel, hence the go-to channel for every plot
• Discriminability can suffer from alignment

Color perception is a complex (neuro-)physiological process.

• Perception of hue depends on adjacent areas

8% of the population suffer from color vision deficiency

• Most frequent manifestation is a red-green deficiency, affecting mostly men

Color can be described, amongst others, by the HSL system

• Hue: the wavelength of the light
• Saturation: how much grey is mixed to the color
• Luminance: how much white is mixed to the color

Luminance

A magnitude channel, suitable for ordered data.

Perception is hampered by contrast effects.

Less than five bins distinguishable.

Saturation

A magnitude channel, suitable for ordered data.

Low accuracy between noncontiguous regions.

Only around three bins distinguishable.

Strong interaction with the size channel.

Saturation and Luminance are rather inefficient magnitude channels, ranging below position, angle, size channels.

Still useful due to versatile application cases.

Remediation: smartly designed color maps, Viridis:

• preserve perceptually-uniformity
• use more than one hue
• account for color vision deficient people
• account for black-and-white printing

Do no use Rainbow color map for sequential data.

• Very effective for categorical data (only second to position cue).
• Between 6-12 discriminable bins.
• Nameable colors are preferred to improve communication.
• High saturation for small areas, low saturation for large areas
• Good resource: https://colorbrewer2.org
• If to many categories: reduce, hierarchically group, introduce “others”

Diverging color maps

• Diverging color maps to depict deviation, e.g. Temperature [\(^{\circ}\)C], log2FC, etc.
• Ternary color maps

Transparency

• Not independently to other color channels. Strong interaction with luminance and saturation, less so with hue
• Limited number of discriminable bins (~2)
• It can be useful to use in combination with other channel to improve popout

• Identity channel
• Applicable to point and line marks, not to area marks
• Strong interference with size channel, and also to hue channel
• In practice limited to a dozen bins

• Magnitude channel
• Useful to communicate periodicity (>2\(\pi\)) and whole ensemble (2\(\pi\))
• Perception of angle is not uniform across the spectrum
  • we can distinguish between 45\(^{\circ}\) and 46\(^{\circ}\) but not between 37\(^{\circ}\) and 39\(^{\circ}\)
  • poor discriminability (~6 bins)
• Tilt is a closely related channel to angle

• Magnitude channel for ordered data.
• Interactions with most other channels
  • if marks too small, shape and color can not be perceived
  • perception of hue and saturation depends on area, hence size
• Size can theoretically encode two values, width and height, which is hardly used since difficult to decipher
• For lines, size is also called linewidth
• Distinguish between scaling the area or the radius of a point.
  • Area makes seems more sensible, but is less accurately perceived.