Freeing Association:
Visual Query Patterns for
Multidimensional Data Dissection
Chris Weaver
School of Computer Science and the Center for Spatial Analysis
University of Oklahoma
The North-East Visualization and Analytics Center
Penn State University
weaver@cs.ou.edu
Analysis in Wonderland
Cheshire Cat: Oh, by the way, if you'd really like
to know, he went that way.
Alice: Who did?
Cheshire Cat: The White Rabbit.
Alice: He did?
Cheshire Cat: He did what?
Alice: Went that way.
Cheshire Cat: Who did?
Alice: The White Rabbit.
Cheshire Cat: What rabbit?
Alice: But didn't you just say - I mean - Oh, dear.
(from “Aliceʼs Adventures in Wonderland” by Charles Lutwidge Dodgson, 1865)
2
Overview
• The Goal
– “Free association” in foraging
– Exploration of relationships between ad hoc categories
• The Approach
– Develop visual interactive techniques for expressing sequences of
multidimensional set queries
– Develop a “lab notebook” for recording, recalling, restoring, and
relating queries and query sequences
• The Results (so far)
– Three reusable high-level design patterns for visual analytics
• Cross-Filtered Views
• Cross-Highlighted Views
• Attribute Relationship Graphs
– A variety of concrete examples with demonstrated utility
3
boat (geospatial cross-filtering)
cell (cascading timeseries – unfinished)
VAST Challenge Tools
evac (cross-highlighted motion traces)
4
wiki (author+word co-occurrences)
Data Dissection
methods for interactively expressing sequences of
multidimensional set queries by visualactively
associating unique data values across multiple views
• Multiple views support selection over sets of unique attribute values in multiple
raw or derived data columns, across one or more tables.
• Attributes are displayed in dimensionally-appropriate view(s) that supports a
binary categorization of values by selection or navigation.
• Users can rapidly toggle dependencies between pairs of views to pose complex
drill-down set queries: effect only those values in view B that co-occur in the
data with the values selected in view A
• Attributes are displayed in a entity-relationship view that shows co-occurrences
(relationships) between values (entities).
• Users can rapidly toggle visibility of attributes and attribute relationships to
dissect the data by slicing in and across data columns
• Analysts can form hypotheses and follow chains of evidence by successive
selection/deselection and filtering/unfiltering of values.
5
boat (cross-filtering)
a visualization of boat landings and interdictions
6
Data Source: VAST 2008 Challenge, Boats Mini-Challenge (synthetic)
Visualization Design: Chris Weaver, Michael Stryker, Ian Turton
Cross-Filtering Queries
• Group (γ) data records into sets for each unique attribute value.
• Filter (φ) each set, keeping records whose attribute values match
those selected in other views.
• Project/visually encode (π) each value and its filtered set.
• Select (σ) values/sets corresponding to brushed glyphs in the view.
7
Cross-Filtering Queries (Boat)
Pre-filtering
Encounters
ee
!ee
G’
#ee
V
"ee
$
G
ed
!ed
G’
#ed
V
"ed
!ec
T’
#ec
V
"ec
ee
T’
!e
G
Casualties
e
$
Date
T
e
ed
number of passengers
number of fatalities
id?
!es
G’
#es
V
"es
$
G
ev
!ev
G’
#ev
V
"ev
Vessel
$
G
er
!er
G’
#er
V
"er
Resolution
$
G
pn
!pn
G’
#pn
V
"pn
Name
Passengers
p
ec
Ship
8
T’
ec
ed
es
er
!p
ed
ee
G
ev
p
ee
$
es
T
Cross-filtering
Encounter
t min <= t <= t max
Grouping
pn
es
ev
er
pn
es
ev
er
pn
Two tables keyed on encounter ID (encounters, passengers)
Seven effective drill-down dimensions
“Variation filtering” of all dimensions on a range of dates
Cross-Filtering...
“I wish you wouldn’t keep appearing and
vanishing so suddenly; you make one quite giddy!”
“All right,” said the Cat; and this time it vanished
quite slowly, beginning with the end of the tail,
and ending with the grin, which remained some
time after the rest of it had gone.
“Well! I’ve often seen a cat without a grin,”
thought Alice; “but a grin without a cat! It’s the
most curious thing I ever saw in all my life!”
(from “Aliceʼs Adventures in Wonderland” by Charles Lutwidge Dodgson, 1865)
9
...Has A Problem
• No automatic value deselection
– Maintain selections to preserve analyst’s ad hoc
groupings
– Avoid filter cascades that terminate at fixed
points/null sets
• But selected values can become invisible!
• Unexpected common case
– Consequence of ‘upstream’ filtering, e.g. C on B
then B on A
– Multiple ways to happen when user “changes
course” analytically
• Problem: An “out of sight, out of mind” effect
?
– Easy to forget items that are selected but invisible
– Leads to misinterpretation when using visible state
of views to remember query clauses
• Solution: Preserve context by assuring visibility
of selected items?
10
evac (cross-highlighting)
a visualization of movements of RDIF-carrying health care workers and visitors
11
Data Source: VAST 2008 Challenge, Evacuation Mini-Challenge (synthetic)
Visualization Design: Chris Weaver and Anthony Robinson
Cross-Highlighting Queries (Evac)
highlight layer only
12
But Crossing Doesn’t Help with...
“Who cares for you?” said Alice, (she had grown to
her full size by this time.) “You’re nothing but a
pack of cards!”
At this, the whole pack rose up into the air, and
came flying down upon her.
(from “Aliceʼs Adventures in Wonderland” by Charles Lutwidge Dodgson, 1865)
13
...Ambiguity of Association
• Example: cross-filter ships, vessels,
resolutions on passengers.
• Which passenger(s) on each
ship? Using each vessel type?
For each kind of resolution?
• Crossed queries are multidimensional and disjunctive.
• Visual states reflect many to
many (to many ...) relationships,
but only show entities.
• Drill into relationships using more
cross-filtering...over the set of all
entity subsets? Tedious!
• Directly visualize co-occurrences?
14
wiki (attribute relationship graph)
a visualization of authors and language over time in a wiki edit history
15
Data Source: VAST 2008 Challenge, Wiki Edit Mini-Challenge (synthetic)
Visualization Design: Chris Weaver, Chi-chun Pan, Don Pellegrino, Prasenjit Mitra
Attribute Relationship Graphs
A multigraph displays attribute values and their cliques — sets
of co-occurrences across pairs of data columns.
Users dynamically filter nodes, edges, and packs (families of
nodes in convex hull wrappers) by selecting particular columns
as well as arbitrary subsets of unique values in those columns.
16
Nominal
Attributes
Temporal
Spatial
KEDS
Design Variations
Auxiliary
Views
17
Post-filter
Retrosheet
Cinegraph
code (event)
name (guest)
name (home team, away team)
name (movie, genre, oscar, person, role)
date (event)
date (visit)
date & time (game)
date (release)
region (countries)
location (hotel, residence)
location (stadium)
-
Numerical cooperative/conflictual weight
Pre-filter
Hotels
-
capacity, attendance, temperature, wind speed box office, rating average, rating count
list (data sources)
-
-
sliders (ratings & roles thresholds)
map (world)
map (Pennsylvania)
map (North America), rich drill-down table
attribute relationship graph
drill-down table
-
movie viewer
1-D heatmap (game count by date)
histogram (rating distribution)
Detail drill-down table, split time series
Nested scatter plot (date vs. weight) 1-D heatmap (visit count by date)
And If We Don’t Remember Our Queries...
Then she began looking about, and noticed
that what could be seen from the old room was
quite common and uninteresting, but that all
the rest was as different as possible. . . . “
They don’t keep this room so tidy as the other,”
thought Alice to herself.
(from “Aliceʼs Adventures in Wonderland” by Charles Lutwidge Dodgson, 1865)
18
...Are We Doomed to Repeat Them?
• Easy to get distracted by new questions
– All three patterns are dangerously “effective” this way
– Chains of interaction reveal many intriguing paths for exploration
• Easy to forget
– past queries
– earlier query clauses expressed in a chain of interactions
• So lots of foraging, not enough sense-making
• Need visual history
• Hmmm...interaction is simple, general, and happens at a
moderate level of interaction abstraction...
19
Idea: Queries to Questions
• “Who, what, where, when” output interface that provides a
summary of query sequences
• Map queries and results into a pseudo natural language
– Designer-specified rule-based text generation
– Parallel illustration with restorable live snapshots of visual state
– Visual highlighting of text that involves attribute values and types
– Inline graphics for enumerating longer sets, à la sparklines
• Cross designs are particularly amenable to such mappings?
– Highly symmetric in form
– Medium level of abstraction
– Simple interactions (select, toggle filter, toggle graph element)
– Most interactions drive interesting, capturable query transitions
20
Harder Than It Seems?
“Let’s consider your age to begin with—how old
are you?” asked the White Queen.
“I’m seven and a half exactly,” said Alice.
“You needn’t say ‘exactly’,” the Queen remarked:
“I can believe it without that. Now I’ll give you
something to believe. I’m just one hundred and
one, five months and a day.”
“I can’t believe that!” said Alice.
(from “Aliceʼs Adventures in Wonderland” by Charles Lutwidge Dodgson, 1865)
21
Thanks!
• Alan M. MacEachren
• Donna Peuquet
• Anthony Robinson
• Students/staff at NEVAC and the GeoVISTA Center
• Collaborators on various CFV/CHV/ARG applications
– Hotels: Deryck Holdsworth & David Fyfe (Penn State/GeoVISTA)
– REMO/Evac: Patrick Laube
– Multiple: The VAST Challenge 2008 team from NEVAC