Applications of Large-Scale Visualization Using Trelliscope

Hafen Consulting, LLC

Purdue University

@hafenstats

Ryan Hafen

Context: Exploratory analysis and statistical model building

Image source: Hadley Wickham

  • Visualization is usually the driver of iteration
  • It is particularly useful when working with a domain expert
  • Interactive visualization often (not always) makes the process more effective
  • The iteration in exploratory analysis necessitates rapid generation of many visualizations
  • A lot of interactive visualization is very customized and time-consuming to create
  • Not every plot is useful so we can't afford to waste a lot of time on any single visualization
  • Just like using a high-level programming language for rapidly trying out ideas with data analysis, we need high-level ways to flexibly but quickly create interactive visualizations

Small Multiples

A series of similar plots, usually each based on a different slice of data, arranged in a grid

"For a wide range of problems in data presentation, small multiples are the best design solution."

Edward Tufte (Envisioning Information)

This idea was formalized and popularized in S/S-PLUS and subsequently R with the trellis and lattice packages

Advantages of Small Multiple Displays

  • Avoid overplotting
  • Work with big or high dimensional data
  • It is often critical to the discovery of a new insight to be able to see multiple things at once
    • Our brains are good at perceiving simple visual features like color or shape or size and they do it amazingly fast without any conscious effort
    • We can tell immediately when a part of an image is different from the rest, without really having to focus on it

Small Multiples

A series of similar plots, usually each based on a different slice of data, arranged in a grid

"For a wide range of problems in data presentation, small multiples are the best design solution."

Edward Tufte (Envisioning Information)

This idea was formalized and popularized in S/S-PLUS and subsequently R with the trellis and lattice packages

Trelliscope:

Interactive Small Multiple Display

  • Small multiple displays are useful when visualizing data in detail
  • But the number of panels in a display can be potentially very large, too large to view all at once

Trelliscope is a general solution that allows small multiple displays to come alive by providing the ability to interactively sort and filter the panels based on summary statistics, cognostics, that capture attributes of interest in the data being plotted

Motivating Example

Gapminder

Suppose we want to understand mortality over time for each country

Observations: 1,704
Variables: 6
$ country   <fctr> Afghanistan, Afghanistan, Afghanistan, Afghanistan, Afgh...
$ continent <fctr> Asia, Asia, Asia, Asia, Asia, Asia, Asia, Asia, Asia, As...
$ year      <int> 1952, 1957, 1962, 1967, 1972, 1977, 1982, 1987, 1992, 199...
$ lifeExp   <dbl> 28.801, 30.332, 31.997, 34.020, 36.088, 38.438, 39.854, 4...
$ pop       <int> 8425333, 9240934, 10267083, 11537966, 13079460, 14880372,...
$ gdpPercap <dbl> 779.4453, 820.8530, 853.1007, 836.1971, 739.9811, 786.113...
glimpse(gapminder)
qplot(year, lifeExp, data = gapminder, color = country, geom = "line")

Yikes! There are a lot of countries...

qplot(year, lifeExp, data = gapminder, color = continent,
  group = country, geom = "line")

Still too much going on...

qplot(year, lifeExp, data = gapminder, color = continent,
  group = country, geom = "line") +
    facet_wrap(~ continent, nrow = 1)

That helped a little...

p <- qplot(year, lifeExp, data = gapminder, color = continent,
  group = country, geom = "line") +
    facet_wrap(~ continent, nrow = 1)
plotly::ggplotly(p)

This helps but there is still too much overplotting...

(and hovering for additional info is too much work and we can only see more info one at a time)

qplot(year, lifeExp, data = gapminder) + theme_bw() +
  facet_wrap(~ country + continent)
qplot(year, lifeExp, data = gapminder) + theme_bw() +
  facet_trelliscope(~ country + continent, nrow = 2, ncol = 7, width = 300)

Note: this and future plots in this presentation are interactive - feel free to explore!

qplot(year, lifeExp, data = gapminder) + theme_bw() +
  facet_trelliscope(~ country + continent,
    nrow = 2, ncol = 7, width = 300, as_plotly = TRUE)

TrelliscopeJS

JavaScript Library

R Package

trelliscopejs-lib
trelliscopejs
  • Built using React
  • Pure JavaScript
  • Interface agnostic
  • htmlwidget interface to trelliscopejs-lib
  • Evolved from CRAN "trelliscope" package (part of DeltaRho project)
devtools::install_github("hafen/trelliscopejs")

Application:

Assessing Fits of Many Models

country_model <- function(df)
  lm(lifeExp ~ year, data = df)

by_country <- gapminder %>%
  group_by(country, continent) %>%
  nest() %>%
  mutate(
    model = map(data, country_model),
    resid_mad = map_dbl(model, function(x) mad(resid(x))))

by_country

Example adapted from "R for Data Science"

# A tibble: 142 × 5
       country continent              data    model resid_mad
        <fctr>    <fctr>            <list>   <list>     <dbl>
1  Afghanistan      Asia <tibble [12 × 4]> <S3: lm> 1.4058780
2      Albania    Europe <tibble [12 × 4]> <S3: lm> 2.2193278
3      Algeria    Africa <tibble [12 × 4]> <S3: lm> 0.7925897
4       Angola    Africa <tibble [12 × 4]> <S3: lm> 1.4903085
5    Argentina  Americas <tibble [12 × 4]> <S3: lm> 0.2376178
6    Australia   Oceania <tibble [12 × 4]> <S3: lm> 0.7934372
7      Austria    Europe <tibble [12 × 4]> <S3: lm> 0.3928605
8      Bahrain      Asia <tibble [12 × 4]> <S3: lm> 1.8201766
9   Bangladesh      Asia <tibble [12 × 4]> <S3: lm> 1.1947475
10     Belgium    Europe <tibble [12 × 4]> <S3: lm> 0.2353342
# ... with 132 more rows

Gapminder Example from "R for Data Science"

  • One row per group
  • Per-group data and models as "list-columns"
country_plot <- function(data, model) {
  figure(xlim = c(1948, 2011),
    ylim = c(10, 95), tools = NULL) %>%
    ly_points(year, lifeExp, data = data, hover = data) %>%
    ly_abline(model)
}

country_plot(by_country$data[[1]],
  by_country$model[[1]])

Plotting the Data and Model Fit for Each Group

We'll use the rbokeh package to make a plot function and apply it to the first row of our data

by_country <- by_country %>%
  mutate(plot = map2_plot(data, model, country_plot))

by_country

Example adapted from "R for Data Science"

# A tibble: 142 × 6
       country continent              data    model resid_mad         plot
        <fctr>    <fctr>            <list>   <list>     <dbl>       <list>
1  Afghanistan      Asia <tibble [12 × 4]> <S3: lm> 1.4058780 <S3: rbokeh>
2      Albania    Europe <tibble [12 × 4]> <S3: lm> 2.2193278 <S3: rbokeh>
3      Algeria    Africa <tibble [12 × 4]> <S3: lm> 0.7925897 <S3: rbokeh>
4       Angola    Africa <tibble [12 × 4]> <S3: lm> 1.4903085 <S3: rbokeh>
5    Argentina  Americas <tibble [12 × 4]> <S3: lm> 0.2376178 <S3: rbokeh>
6    Australia   Oceania <tibble [12 × 4]> <S3: lm> 0.7934372 <S3: rbokeh>
7      Austria    Europe <tibble [12 × 4]> <S3: lm> 0.3928605 <S3: rbokeh>
8      Bahrain      Asia <tibble [12 × 4]> <S3: lm> 1.8201766 <S3: rbokeh>
9   Bangladesh      Asia <tibble [12 × 4]> <S3: lm> 1.1947475 <S3: rbokeh>
10     Belgium    Europe <tibble [12 × 4]> <S3: lm> 0.2353342 <S3: rbokeh>
# ... with 132 more rows

Apply This Function to Every Row

A plot for each model

by_country %>%
  trelliscope(name = "by_country_lm", nrow = 2, ncol = 4)

Application:

Images as Panels

(Database of Visualizations)

pokemon <- read_csv("http://bit.ly/plot_pokemon") %>%
  mutate_at(vars(matches("_id$")), as.character) %>%
  mutate(panel = img_panel(url_image))

pokemon
trelliscope(pokemon, name = "pokemon", nrow = 3, ncol = 6,
  state = list(labels = c("pokemon", "pokedex")))
read_csv("http://bit.ly/trs-mri") %>%
  mutate(img = img_panel(img)) %>%
  trelliscope("brain_MRI", nrow = 2, ncol = 5)

A Larger Dataset: Growth Trajectories of >2k Children

(offline demo)

Case Study:

Exploring 44.5 Million Live Births in Brazil

The Data

  • Publicly available 
  • ​Data from 2001 to 2015
  • 44.5 million birth records
  • Analyzed in memory
Observations: 44,509,207
Variables: 27
$ dn_number         <chr> "05558306", "05559894", "05559900", "10660701", "...
$ birth_place       <fct> Hospital, Hospital, Hospital, Hospital, Hospital,...
$ health_estbl_code <chr> "0000001", "0000009", "0000009", "0000006", "0001...
$ birth_muni_code   <int> 110009, 110002, 110002, 110002, 120070, 120070, 1...
$ m_age_yrs         <dbl> 25, 15, 35, 17, 31, 23, 24, 16, 15, 19, 19, 19, 2...
$ marital_status    <fct> Single, Single, Married, Single, Single, Single, ...
$ m_educ            <fct> 4 to 7 years, 1 to 3 years, 8 to 11 years, 1 to 3...
$ occ_code          <chr> "00800", NA, NA, NA, "31000", "00800", "00800", N...
$ n_live_child      <dbl> 1, NA, 2, 1, 1, 1, 1, NA, NA, 2, 3, NA, 4, NA, NA...
$ n_dead_child      <dbl> 0, NA, NA, NA, 0, 0, 0, NA, NA, NA, NA, NA, 1, NA...
$ m_muni_code       <int> 120040, 120010, 120025, 120040, 120070, 120070, 1...
$ gest_weeks        <fct> 37-41 weeks, 37-41 weeks, 37-41 weeks, 37-41 week...
$ preg_type         <fct> Singleton, Singleton, Singleton, Singleton, Singl...
$ deliv_type        <fct> Vaginal, Vaginal, Cesarean, Vaginal, Vaginal, Vag...
$ n_prenatal_visit  <fct> 1 - 3, 4 - 6, 7+, 1 - 3, 4 - 6, 1 - 3, 7+, 7+, 7+...
$ birth_date        <date> 2001-02-20, 2001-03-30, 2001-06-07, 2001-12-05, ...
$ sex               <fct> Male, Male, Male, Female, Female, Female, Male, F...
$ apgar1            <int> 8, 8, 8, 9, 7, 6, NA, NA, 3, 5, 5, 5, 5, 5, NA, 7...
$ apgar5            <int> 10, 10, 10, 10, 8, 8, NA, 8, 7, 8, 8, 9, 9, 10, 8...
$ race              <fct> White, White, White, White, White, White, Multira...
$ brthwt_g          <dbl> 3800, 3100, 3300, 3200, 3600, 3700, 3750, 3500, 3...
$ cong_anom         <fct> No, No, No, No, NA, NA, No, No, No, No, No, No, N...
$ cong_icd10        <chr> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, NA, N...
$ birth_year        <int> 2001, 2001, 2001, 2001, 2001, 2001, 2001, 2001, 2...
$ m_state_code      <chr> "AC", "AC", "AC", "AC", "AC", "AC", "AC", "AC", "...
$ birth_state_code  <chr> "RO", "RO", "RO", "RO", "AC", "AC", "AC", "AC", "...
$ m_age_bin         <fct> 20-29, 10-19, 30-39, 10-19, 30-39, 20-29, 20-29, ...

Low Birth Weight Over Time by State

Tangent: geofacet

Low Birth Weight by Municipality

ggplot(by_muni_lbwt_time, aes(birth_year, pct_low_bwt)) +
  geom_point(size = 3, alpha = 0.6) +
  geom_line(stat = "smooth", method = rlm,
  color = "blue", size = 1, alpha = 0.5) +
  theme_bw() +
  labs(y = "Percent Low Birth Weight", x = "Year") +
  facet_trelliscope(~ state_name + muni_name,
    nrow = 2, ncol = 4, width = 400, height = 400,
    name = "pct_low_bwt_muni",
    desc = "percent low birth weight yearly by municipality")

(offline demo - for now)

Method of Delivery

Scaling Trelliscope

Just because you can't look at all panels in a display doesn't mean it isn't useful or practical to make a large display - it's in fact beneficial because you get an unprecedented level of detail in your displays, and every corner of your data can be conceptually viewed

One insight is all you need for a display to serve a purpose (provided it is quick to create)

We used the previous implementation of Trelliscope to visualize millions of subsets of terabytes of data

What is needed to scale in the Tidyverse?

SparklyR is the natural solution

But we need a few things...

  • SparklyR support for list-columns (nested data frames and arbitrary R objects)
  • Fast random access to rows of a SparklyR data frame
  • A TrelliscopeJS deferred panel rendering scheme (render on-the-fly rather than all panels up front)

Ongoing Work

  • Trelliscope
    • Deferred panels for very large displays
    • Automatic determination of how "interesting" a given partitioning will be based on what is being plotted
    • When axes are "same", only show axes on plot margins instead of every panel (underway for ggplot2)
  • trelliscopejs-lib
    • More visual filters for cognostics (dates, geographic, bivariate relationships, etc.)
    • Bookmarkable / sharable state
    • View multiple panels from different displays on same conditioning side-by-side

For More Information

install.packages(c("tidyverse", "gapminder", "rbokeh", "plotly"))
devtools::install_github("hafen/trelliscopejs")

library(tidyverse)
library(gapminder)
library(rbokeh)
library(trelliscopejs)

Most examples in this talk are reproducible after installing and loading the following packages:

Trelliscope Displays as Apps

library(shiny)
library(ggplot2)
library(gapminder)

server <- function(input, output) {
  output$countryPlot <- renderPlot({
    qplot(year, lifeExp,
      data = subset(gapminder,
        country == input$country)) +
      xlim(1948, 2011) + ylim(10, 95) +
      theme_bw()
  })
}

choices <- sort(unique(gapminder$country))
ui <- fluidPage(
  titlePanel("Gampinder Life Expectancy"),
  sidebarLayout(
    sidebarPanel(
      selectInput("country",
        label = "Select country: ",
        choices = choices,
        selected = "Afghanistan")
    ),
    mainPanel(
      plotOutput("countryPlot",
        height = "500px")
    )
  )
)

runApp(list(ui = ui, server = server))

Trelliscope Displays as Apps

If you have an app that has multiple inputs and produces a plot output, the idea is simply to enumerate all possible inputs as rows of a data frame and add the plot that corresponds to these parameters as column and plot it

Trelliscope displays are most useful as exploratory plots to guide the data scientist (because they can be created rapidly)

However, in many cases Trelliscope displays can be used as interactive applications for end-users, domain experts, etc. with the bonus that they are much easier to create than a custom app

From ggplot2 Faceting to Trelliscope

Turning a ggplot2 faceted display into a Trelliscope display is as easy as changing:

to:

facet_wrap()

or:

facet_grid()
facet_trelliscope()

TrelliscopeJS in the Tidyverse

  • Create a data frame with one row per group, typically using Tidyverse group_by() and nest() operations
  • Add a column of plots
    • TrelliscopeJS provides purrr map functions map_plot(), map2_plot(), pmap_plot() that you can use to create these
    • You can use any graphics system to create the plot objects (ggplot2, htmlwidgets, lattice)
  • Optionally add more columns to the data frame that will be used as cognostics - metrics with which you can interact with the panels
    • All atomic columns will be automatically used as cognostics
    • Map functions map_cog(), map2_cog(), pmap_cog() can be used for convenience to create columns of cognostics
  • Simply pass the data frame in to trelliscope()

With plots as columns, TrelliscopeJS provides nearly effortless detailed, flexible, interactive visualization in the Tidyverse

library(visNetwork)

nnodes <- 100
nnedges <- 1000

nodes <- data.frame(
  id = 1:nnodes,
  label = 1:nnodes, value = rep(1, nnodes))
edges <- data.frame(
  from = sample(1:nnodes, nnedges, replace = T),
  to = sample(1:nnodes, nnedges, replace = T)) %>%
    group_by(from, to) %>%
    summarise(value = n())

network_plot <- function(id, hide_select = TRUE) {
  style <- ifelse(hide_select,
    "visibility: hidden; position: absolute", "")

  visNetwork(nodes, edges) %>%
    visIgraphLayout(layout = "layout_in_circle") %>%
    visNodes(fixed = TRUE, scaling =
      list(min = 20, max = 50,
        label = list(min = 35, max = 70,
          drawThreshold = 1, maxVisible = 100))) %>%
    visEdges(scaling = list(min = 5, max = 30)) %>%
    visOptions(highlightNearest = list(enabled = TRUE, degree = 0,
      hideColor = "rgba(200,200,200,0.2)"),
      nodesIdSelection = list(selected = as.character(id), style = style))
}

network_plot(1, hide_select = FALSE)

Network Vis with visNetwork htmlwidget

nodedat <- edges %>%
  group_by(from) %>%
  summarise(n_nodes = n(), tot_conns = sum(value)) %>%
  rename(id = from) %>%
  arrange(-n_nodes) %>%
  mutate(panel = map_plot(id, network_plot))

nodedat
# A tibble: 100 × 4
      id n_nodes tot_conns            panel
   <int>   <int>     <int>           <list>
1     58      17        19 <S3: visNetwork>
2     45      16        17 <S3: visNetwork>
3      9      15        18 <S3: visNetwork>
4     31      15        16 <S3: visNetwork>
5     14      14        15 <S3: visNetwork>
6     42      14        15 <S3: visNetwork>
7     90      14        14 <S3: visNetwork>
8     21      13        14 <S3: visNetwork>
9     37      13        14 <S3: visNetwork>
10    43      13        13 <S3: visNetwork>
# ... with 90 more rows

Trelliscope display with one panel per node

We create a one-row-per-node data frame with number of nodes connected to and total number of connections as cognostics and add a plot panel column

nodedat %>%
  arrange(-n_nodes) %>%
  trelliscope(name = "connections", nrow = 2, ncol = 4)

Applications of Large-Scale Visualization Using Trelliscope

By Ryan Hafen

Applications of Large-Scale Visualization Using Trelliscope

  • 1,014
Loading comments...

More from Ryan Hafen