Visualization for Machine Learning

What is Information Visualization? Why Use It?

“The use of computer-supported, interactive, visual representations of abstract data to amplify cognition.”

Key Concepts

Computer-Based
Visual Representation
Abstract Data
Interactive
Amplify Cognition

Abstract Data

Data with no obvious/natural visual representation

Abstract Data

Data with no obvious/natural visual representation

Interactive

Users can change what is visualized and how it is visualized.

Amplify Cognition

Solve problems with data with less effort, in a shorter time, and more accurately.
… or even be able to do things it would be impossible to do without a computer and a graphical representation.

Cognitive artifacts: tools that help us think!

Try to multiply 34 x 72 using exclusively your mind …
… now do it again using pen and paper.

Why is it easier?

… because we can store intermediary results in the paper rather than keeping the information in mind. That is, part of the memory is in the world rather than in your head.

Let’s play the “game of 15” …

The “pieces” for the game are the nine digits: 1, 2, 3, 4, 5, 6, 7, 8, 9. Each player takes a digit in turn. Once a digit is taken, it cannot be used by the other player. The first player to get three digits that sum to 15 wins.
Here is a sample game: Player A takes 8. Player B takes 2. Then A takes 4, and B takes 3. A takes 5.
Question 1: Suppose you are now to step in and play for B. What move would you make?

Let’s play a different game: tic-tac-toe

Players alternately place an O or a X in one of nine spaces arranged in a rectangular array. Once a space has been taken, it cannot be changed by either player. The first player to get three symbols in a straight line wins. Suppose player A is X and B is O, and the game has reached the state on the right.

Question 2: Suppose you are now to step in and play an O for B. What move would you make?

Problem Isomorphs Herbert Simon

The two problems are equivalent!

Why use visualization?

Explanatory visualization
Exploratory visualization
Confirmatory visualization

Great Explanatory Visualizations

NYT: https://flowingdata.com/tag/new-york-times/
Washington Post: http://postgraphics.tumblr.com/
Gregor Aisch: https://driven-by-data.net/
Nicky Case/Explorable Explanations: http://explorabl.es/
Polygraph: http://polygraph.cool/ & https://pudding.cool/
ProPublica: https://www.propublica.org/

Why use a graphical representation?

Large parts of our brain are devoted to spatial processing

Why use a computer to visualize data?

Why use interaction?

Each visualization can only answer a subset of questions.
With interaction the user can change what is visualized and how to answer a multitude of questions.
Also one cannot visualize everything at once.

How do you assess the quality of a visualization?

Isn’t it subjective? Some people like A, whereas some others like B.

Some visual representations are better than others at solving particular problems …

Digression: Graphical Perception

Graphical Perception Experiment

Graphical Perception Results

Designing effective visualizations requires

Knowing the design space
Being able to compare the solutions
… in turn comparing the solutions requires understanding human perception.

Data Types

The first ingredient in effective visualization is the input data. Data values can represent different forms of measurement.
What kinds of comparisons do those measurements support?
What kinds of visual encodings then support those comparisons?

Nominal (N) or Categorical (C)

Nominal data — also called categorical data — consist of category names.
With nominal data we can compare the equality of values: is value A the same or different than value B? (A = B), supporting statements like “A is equal to B” or “A is not equal to B”.
When visualizing nominal data we should readily perceive if values are the same or different: position, color hue (blue, red, green, etc.), and shape are all reasonable options.

Ordinal (O)

Ordinal data consist of values that have a specific ordering.
With ordinal data we can compare the rank-ordering of values: does value A come before or after value B? (A < B), supporting statements like “A is less than B” or “A is greater than B”.
When visualizing ordinal data, we should perceive a sense of rank-order. Position, size, or color value (brightness) might be appropriate, whereas color hue (which is not perceptually ordered) would be less appropriate.

Quantitative (Q)

With quantitative data we can measure numerical differences among values.
There are multiple sub-types of quantitative data:
- For interval data we can measure the distance between points: (A - B).
- For ratio data we can also measure proportions or scale factors: (A / B).
Quantitative values can be visualized using position, size, or color value, among other channels. An axis with a zero baseline is essential for proportional comparisons of ratio values, but can be safely omitted for interval comparisons.

Temporal (T)

Temporal values measure time points or intervals. This type is a special case of quantitative values (timestamps) with rich semantics and conventions (i.e., the Gregorian calendar).
Example temporal values include date strings such as “2019-01-04” and “Jan 04 2019”, as well as standardized date-times such as the ISO date-time format: “2019-01-04T17:50:35.643Z”. There are no temporal values in our global development dataset above, as the year field is encoded as an integer.

Spatial (S)

Data that can be shown in a map
Also known as geospatial data, refers to information that identifies the geographic location and characteristics of natural or constructed features and boundaries on the Earth. https://atlan.com/spatial-data/

Data Types Summary

These data types are not mutually exclusive, but rather form a hierarchy: ordinal data support nominal (equality) comparisons, while quantitative data support ordinal (rank-order) comparisons.
Moreover, these data types do not provide a fixed categorization. For example, just because a data field is represented using a number doesn’t mean we have to treat it as a quantitative type! We might interpret a set of ages (10 years old, 20 years old, etc.) as nominal (underage or overage), ordinal (grouped by year), or quantitative (calculate average age).

Fundamental Charts

Widely adopted, effective, useful.
Solve very large percentage of vis problems.
Training ground for more sophisticated graphs.

Bar Chart

Visualize how a quantity distributes across a set of categories.

Line Chart

Visualize how a quantity changes in relation to another quantity (typically time).

Scatter Plot

Visualize how a quantity relate to another quantity.

Matrix

Visualize how a quantity distributes across two categories.

Symbol Map

Visualize how a quantity distributes across two spatial coordinates.

Fundamental Graphs Summary

Scatter Plots + Faceting (without)

Scatter Plots + Faceting (with)

Tidy Data

Goal: organizing data to make visualization easier

link to paper

Tidy Data

Tidy Data: Definition

In tidy data:

Each variable forms a column.
Each observation forms a row.
Each type of observational unit forms a table.

Tidy Data: Example #1

Tidy Data

Tidy Data: Example #2

Tidy Data

Graphical Encoding

Every visualization can be described in terms of:

its basic graphical components
mapping strategy between data and graphics
more precisely, a set of mappings between:
- data items — visual marks
- data attributes — visual channels

Graphical Marks

Visual Encoding Channels

Visualization Decoding

Marks — Data Items
Channels — Data Attributes

Examples

NYT link

Expressiveness Principle

The visual representation should express the type of information that exists in the data.

Ordered data should not appear as unordered.
Unordered data should not appear as ordered.

Effectiveness Principle

Relevance of information should match the effectiveness of the channels used.

Represent important information with more effective channels

Effectiveness Effect

Summary

Visual Encoding/Decoding
Graphical Marks and Channels
Expressiveness and Effectiveness
Channels Appropriateness and Ranking
Evaluation and Design
Contextual Components
- Labels, legends and annotations
- Axes, grids and trend lines

Neo: Interactive Confusion Matrices

Great example of research in VisML!

video link