About Me

This Discussion

1. Talk reviews

2. Literature Review: 3 taxonomies for interactive graphics

3. A small example of interactive graphics to do forecasting @ Amazon

Hafen: Trelliscope

• Combines small multiples with interactive sorting & filtering of (> 1000s) of panels

Schep: iheatmapr

• Heatmaps that are complex (multiple graphs laid out modularly) and interactive.

Robinson: Interactive Vis @ Stack Overflow

• Network graph of languages, time series graphs of languages, and maps of developers

Interactive Visualization Taxomonies

Three very nice review papers: (1) Stats, (2) HCI, (3) InfoVis

1. 1996. A. Buja, D. Cook, and D. Swayne, “Interactive High-Dimensional Data Visualization”, Journal of Computational and Graphical Statistics
   • Statistics literature
2. 1996. B. Shneiderman, “The Eyes Have It: A Task by Data Type Taxonomy for Information Visualizations”, IEEE (Proc. Symp. on Visual Languages)
   • Human-Computer Interaction HCI (CS) literature
3. 2007. J. Heer, G. Robertson, “Animated Transitions in Statistical Data Graphics”, IEEE Trans. Visualization & Comp. Graphics (Proc. InfoVis)
   • InfoVis (CS) literature

Buja, Cook, & Swayne (1996) Interactive High-Dimensional Data Visualization

Propose a hierarchy of interactive graphics, focusing on

• manipulating (vs. rendering)

• multivariate data (vs. univariate or other types)

• for exploratory analysis (vs. explanatory)

Quote: “The gain resulting from animation and real-time control is not incremental but a quantum leap”

Buja/Cook/Swayne Taxonomy (Tasks / Tools)

Buja/Cook/Swayne Taxonomy (Tasks / Tools)

Buja/Cook/Swayne Taxonomy (Tasks / Tools)

Buja/Cook/Swayne Taxonomy (Tasks / Tools)

Buja/Cook/Swayne Taxonomy (Tasks / Tools)

Shneiderman (1996)

Type by Task Taxonomy (TTT): 7 data types x 7 interactive tasks

Quote: “Visual Information Seeking Mantra: Overview first, zoom and filter, then details-on-demand”

Shneiderman (1996)

One could map elements of the Buja/Cook/Swayne taxonomy to this one…

Quote: “Visual Information Seeking Mantra: Overview first, zoom and filter, then details-on-demand”

Shneiderman (1996)

One could map elements of the Buja/Cook/Swayne taxonomy to this one…

Quote: “Visual Information Seeking Mantra: Overview first, zoom and filter, then details-on-demand”

Heer & Robertson (2007)

Animated Transitions in Statistical Data Graphics

• Taxonomy of transition types
  • View Transformation
  • Substrate Transformation
  • Filtering
  • Ordering
  • Timestep
  • Visualization Change
  • Data Schema Change

Quote: “The challenge of designing animations is to visually interpolate the syntactic features [visual elements] such that semantic changes [meaning of the data] are most effectively communicated.”

Demand Forecasting @

At Amazon, the Supply Chain Optimization Technology (SCOT) organization handles forecasting, buying, and placement of retail items.

• We forecast future demand for tens of millions of products, across many geographic regions
• We publish our forecasts on a daily basis
• We produce forecasts for a range of
  • time grains {daily, weekly, monthly} and
  • horizons starting {immediately, one week from now, one month from now, …}

A major need: visualize and understand time series data to develop models (i.e. iterate between “visualize” and “model” here in the figure below)

[image: “R for Data Science” by Grolemund and Wickham]

Time Series Exploratory Data Analysis

Static ggplot of product sales time series. Not a bad start…

My old workflow

### 1 ###  draw a single plot containing as much info as possible (within reason)
draw_canonical_plot <- function(product) {
  # spend a long time (hours or days) creating a plot for a single instance (product)
  # prototype by trying on a few randomly sampled instances
}

### 2 ### compute a summary statistic of interest
mean_demand <- mean(product$demand)

### 3 ### take a stratified sample of size, say, n = 500:
stratified_sample <- ...

### 4 ### create the plots and collect into one .pdf file:
pdf(file = "plots_stratified_mean_demand.pdf", width = 10, height = 6)
for (i in 1:500) {
  draw_canonical_plot(product[stratified_sample[i]])
}
dev.off()

### 5 ### spend hours poring over the plots & share (via email attachment)

Using Trelliscope: Nice!

Trelliscope offers a great speedup:

1. Sorting by a summary statistic is easy

2. Stratified sampling is possible (by creating dummies for quintiles and then sorting, for instance…)

3. Linked brushing allows for comparisons; query a single variable and see the result across panels.

What about a slider for the smoothing spline?

[quick live demo in R]

1. R’s `manipulate’

2. R plotly (better)

3. a pure javascript solution: probably best, how hard to code/achieve?

A final thought:

Question: What types of interactive visualization:

1. Increase productivity substantially?

2. Increase productivity by a “quantum leap”?

3. Allow for discoveries that were truly not possible using static graphics?

Thanks for three great talks!

1. Applications of Interactive Exploration of Large Multi-Panel Faceted Displays Using Trelliscope
   • Ryan Hafen, Purdue University
2. Extracting Insights from Genomic Data via Complex Interactive Heatmaps
   • Alicia Schep, Stanford School of Medicine
3. Visualizing the Changing Landscape of Software Developers
   • David Robinson, StackOverflow.com

Also a big thanks to Gabe Becker (Genentech Research) for organizing the session around a great topic.

Also, thanks to Carson Sievert for helpful discussions and a pointer to his Ph.D. thesis, which discusses recent developments in interactive graphics and provides historical references:

https://github.com/cpsievert/phd-thesis

Info

• Kenny Shirley

• Amazon Supply Chain Optimization Technology (SCOT)

  • YouTube Video: https://www.youtube.com/watch?v=ncwsr1Of6Cw
  • Website: https://www.amazon.jobs/en/teams/supply-chain-optimization-technologies