LargeScale Cognition: The Psychology of Informavores Dept' of Psychology, Stanford University 14 Mar

1
Large-Scale CognitionThe Psychology of
InformavoresDept. of Psychology, Stanford
University14 March 2001

• Stuart Card
• Xerox PARC

Work supported in part by the Office of Naval
Research
2
Aim of this Talk

• Broad view
• Emphasize scale change in information environment
• Sample of psychological investigations
• Sample of applications

3
Contributors

• Andrew Faurling
• Ruth Rosenholtz
• Ed Chi
• Jim Pitkow
• Jeff Heer
• Chris Olston

• Peter Pirolli
• Mija Van Der Wege
• Paul Whitmore
• Jenea Boshart
• Julie Morrison
• Pam Schraedly
• Rob Reeder
• Allison Woodruff

4
Technology Underpinnings
5
The Information Big Bang
AMOUNT OF INFORMATIONPRODUCED IN THE WORLD
0.6 EB/Yr 50/Yr
2 EB/Yr 50/Yr
6
Comparisons

• 1999 2009
• (EB) (EB)
• Unique Info 0.6 36
• Store in earths populationmemory in 1 yr 0.1
  0.1
• Record all words in alllives 3.6 3.6

7
Communications

• One Net accessible everywhere
• 25M Pages/Day
• 2X 10X more/yr for Seated Home User
• Indexed
• Automatable

???
1 Gb
SAN/backpanels
LAN
1 Mb
???
WAN
ISDN
POTS _at_ 17/year
1 Kb
POTS
Source G. Bell
8
Humans are informavores

• Informavores Organisms that hunger for
  information about the world and themselves
  (George Miller, 1983)
• Humans seek, gather, share, and consume
  information in order to adapt
• Now taking this to new scaleLarge-Scale
  Cognition
• Now that youve got it, how do you every find
  anything?
• How do you make sense of it?

9
Why Interesting to Study

• Important problem
• Information overload
• But potentials for improvement
• High economic costs and benefits
• Web Instrumented
• Can study what was impossible before
• Inherently psychological
• Attention
• Perception
• Memory
• Problem solving

10
Pressures of the information environment

A wealth of informationcreates a poverty of
attention and a need to allocate it efficiently

Herbert Simon
Human attention is central,not precision vs.
recall.
11
Example 1 Field Study Business Intelligence

• Professional Technology Analyst
• TASK Write monthly newsletter
• METHOD Group scans 600 magazines Copy out
  articles Analyst whittles down pile

12
Analyst Workflow
Select File
File
(a)
New Article
Marked Mag.
New Mags 2800 pages
ArticleCopies
ArticleCopies
ArticleCopies
Copy
LIBRARY
Transport
Transport
Transport
Transport
ANALYST
(c)
Marked Mag.
Maga-zines
ArticleCopies
Shuffle Dot
Write Article
New Article
Scan Mark
Project Pile
Project Pile
1 Pile
Sort to Pile
File
Cleanup
Discard
13
Cascaded Filters Concentrates Information
at Low Cost
50 Mags 2800 page/mo(210 hr/mo)
12
1
Marked Arts.1288 page/mo(97 hr/mo)
Project Pile3000 pages(255 hr)
Writing Pile250 pages(19 hr)
Article
Dot Sort
Write article
Scan Mark
To Library(Copy article)
14
Concentration of Information
15
Sensemaking Based on Schemata
16
Ecological Paradigm
17
Ecological Approach

• Human-computer interaction is adaptive to the
  extent

Net Knowledge Gained
MAXIMIZE
Costs of Interaction
18
Information Foraging Theory

• Take concept of informavores seriously
• Key ideas
• Cost structure of information and economics of
  attention
• Information scent. Local cues used to explore
  and search information spaces
• Analysis on two levels
• adaptionist level Rational analysis and
• proximal mechanisms.
• Key implications for machine aids
• Machines that
• (1) Predict Degree-of-Interest over information
  field
• (2) Use information visualization to aid external
  cognition

19
Time scales of analysis
Newell
Time scale (s)
Psychological domain
107 106 105

• SOCIAL

104 103 102

• RATIONAL(Adaptive)

• Task

101 100 10-1

• Unit task
• Operations
• Visual attention

• COGNITIVE(Proximal Mechanisms)

20
E.G. Housefly Foraging For Food
Adaptionist Level Fly near my
sandwich Proximal Mechanisms
Bell
21
Information patches
E.g. Desk Piles, Alta Vista Search List Unlike
animals foraging for food, humans can do patch
construction
22
People are information rate maximizers
Benefits/Costs
23
When to Switch PatchesRandomly-Ordered Prey
Cumulative gain g(tW)
R
R2
R1
t1
t
t2
tB
tW
Between-patch time
Within-patch time
24
Charnovs Marginal Value Theorem
Max gain when slope of within-path gain g
average gain R (tangent in diagram)
Gain
R
g(tW)
Within-patch time
Between-patch time
tB
t
25
Between-Patch Enrichment
Gain
R2
R1
g(tW)
Within-patch time
Between-patch time
tB1
t1
tB2
t2
enrichment
Example arrange physical office efficiently
26
Within-Patch Enrichment
Example Better filtering of search hits
enrichment
g1(tW)
Within-patch time
Between-patch time
27
Summary of Field Study

• Sensitive to cost structure
• Evidence for maximizing Rinfo gain/cost
• Information patches
• Between-patch enrichment (physically)
• Within-patch enrichment (filtering)

28
Example 2. Spiral Calendar
29
Direct Walk Interactions
Display2
Display3
Display1
Etc
Click,Gesture, Etc
Click,Gesture, Etc
Click,Gesture, Etc
Examples WWW, Mac Finder, HyperCard
30
Cost of Knowledge Characteristic Function
Gain in Knowledge
Cost Time
31
COKCF Spiral Calendar
107
Spiral Calendar
106
Calendar Manager
105
104
Items accessed within cost
103
102
101
100
0
20
40
60
80
100
120
COST (s)
32
COKCF Spiral Calendar
33
Summary of Spiral Calendar

• Can measure cost structure
• Cost of Knowledge Characteristic Function

34
Example 3. Scatter/Gather

• supports exploration/browsing of very large
  full-text collections ( 1,000,000)
• creates clusters of content-related documents
• presents users with overviews of cluster contents
• allows user to navigate through clusters and
  overviews

35
Marti Hearst
36
Scatter/Gather task
Display Titles Window
Scatter/Gather Window
Law
Nat. Lang.
World News
Robots
AI
Expert Sys
CS
Planning
Medicine
Bayes. Nets
37
information scent
new
cell
Information Need
medical
patient
Text snippet
treatments
dose
procedures
beam

• Spreading activation
• Derived from models of human memory
• Activation reflects likelihood of relevance given
  past history and current context
• Approximates Bayesian network

38
spreading activation networks(for modeling
scent)
Document corpus
Word statistics
Spreading activation network
39
interface provides good scent of underlying
document clustering
Perceived by model
Identified by computer
40
Summary of Scatter Gather

• Cost structure
• Information Patches (clustered docs)
• Maximizing info gain/cost
• Patch enrichment vs. exploitation
• Spreading activation to model semantic content
• Information scent on direct walk interface
  predicts behavior

41
Example 4Web Study - Protocol Analysis

• Protocol Structure
• URL
• Observed Actions and Transcript
• Protocol Analysis

42
Study

• WWW Task Bank Survey (N 2188)
• 6 Find information tasks, e.g.,
• You are Chair of Comedic events for Louisiana
  State University in Baton Rouge. Your computer
  has crashed and you have lost several
  advertisements for upcoming events. You know that
  the Second City tour is coming to your theatre in
  the spring, but you do not know the precise date.
  Find the date the comedy troupe is playing on
  your campus. Also find a photograph of the group
  to put on the advertisement.
• 12 Stanford University students
• 2 tasks (CITY, ANTZ) analyzed for 4 participants

43
Video Data
Web Logger
Question
Internet Explorer
44
Instrumentation
45
WebLogger Event File
46
Protocol
47
RobReeder
48
(No Transcript)
49
Analysis - Information structure

• Web sites
• Portals
• Search engines
• Pages
• Website home page
• Search engine page
• Hitlist page
• Content elements

50
Problem space structure

• URL
• Link
• Keyword
• Visual Search

51
Web Behavior Graph
Mija Van Der Wege
52
Web Behavior Graph
Link Problem Space
URL Problem Space
Keyword Problem Space
Visual Search Problem Space
53
Web Behavior Graph
Execution of Operator
Return to Previous State
54
Web Behavior Graph
123 Posters
Yahoo
55
Web Behavior Graph
No Scent
Low Scent
Medium Scent
High Scent
56
Web Behavior Graphs (WBGs)
ANTZ
S1
S6
S7
S10
CITY
S1
S6
S7
S10

57
18. People Switch When Information Scent
Gets Low
Patch-leaving policy Leave Web site when
information scent goes below some threshold
58
Phase shifts in search regime due to information
scent
q prob of going down wrong path. Small change
in q, big change in nodesexamined. (Working on
now)
Nodes
Examined
Number of Levels (D) (z 10)
59
Summary of Web Search

• Cost structure
• Patches of sites, pages, content, links
• Maximize info gain/cost by shifting search
  branches and problem spaces
• Basic model of web search
• Follow information scent until weak.
• Shift problem spaces if impasse.
• Small shift in scent can be magnified to large
  cost of search if near phase boundary.
• Direct Walk analysis
• Applies to device controls, software, information
  design

60
SUMMARY - ADAPTIONIST LEVEL

• Information has a cost structure
• Can articulate with, e.g., COKCF
• People are sensitive the cost structure
• Seek to maximize info gained/cost
• Can predict behavior just by analyzing cost
  structure and information scent strength
• Information often in patches
• Between patch costs
• Within patch costs
• Exploration vs exploitation trade-offs
• Opportunity costs
• People can shape information environment as well
  as search, e.g., by enrichment
• Between-patch enrichment
• Within-patch enrichment

61
SUMMARY - PROXIMAL MECHANISMS

• People work in multiple problem spaces determined
  by system, shift on impasse.
• Information ScentBasic method is to follow,
  backtrack on impasse or shift problem space.
• Dominating effect Phase changes
• Small change in probability of going down wrong
  path has large, qualitative effect on search

62
APPLICATION 1Scent Web Site Usability

• Improve website design directly by means of
  improving information scent

63
APPLICATION 2 Web User Flow by Information Scent
(WUFIS) Pete Pirolli Ed Chi
Web site

• Use scent models to simulate user

Web Page content
links
User Information goal
Web user flow simulation
Predicted paths
64
User Flow Model
User need (vector of goal concepts)
65
(No Transcript)
66
APPLICATION 3Enhanced Thumbnails
Improve scent by computed Degree-of-Interest
Visualization

• Text summaries
• Lots of abstract, semantic information
• Image summaries (plain thumbnails)
• Layout, genre information
• Enhanced thumbnails
• Combine features of text thumbnails

67
System
Allison Woodruff

• Preprocessor modifies HTML
• e.g., increase size of text, modify color of text
• Renderer creates scaled image of page
• Postprocessor transforms image
• e.g., apply color wash, add text callouts

68
System

• Preprocessor modifies HTML
• e.g., increase size of text, modify color of text
• Renderer creates scaled image of page
• Postprocessor transforms image
• e.g., apply color wash, add text callouts

69
System

• Preprocessor modifies HTML
• e.g., increase size of text, modify color of text
• Renderer creates scaled image of page
• Postprocessor transforms image
• e.g., apply color wash, add text callouts

70
Examples

• Emphasize text that is relevant to query
• Text callouts
• Enlarge text that might be helpful in assessing
  page
• Enlarge headers

71
Results
Note N 12
72
APPLICATION 4Degree of Interest Trees Stu Card
David Nation

• Increase info gain/cost by computing DOI Info
  Vis.
• Automatic visual patch enrichment
• Maintain contextual orientation

73
APPLICATION 5 Web Forager

• Create patches of pages (WebBooks)
• Create workspace patch
• Enhance COKCF - large space gestures

N
74
SUMMARY - Computer Aids

• Degree-of-Interest Visualization
• Does automatic filtering (within-patch
  enrichment)
• Support working memory and context
• Enhance Information Scent
• Position on good side of Phase Changes
• Next ACT-R models of search, Models of visual
  attention, Information Crystallization

75
(No Transcript)
76
Most Information Not Paper
Only 0.003 of data (by size) is print. 86
is magnetic.
77
Digital Information Taking Over
Paper 2/Yr Film 4/Yr
Non-Digital
InformationProduction
Digital
Optical 70/Yr Magnetic 55/Yr
78
Democratization of Data

• Individuals create
• 80 of original paper documents
• 99 of original film documents
• Individuals have
• 55 of disk drive memory

79
What Were Trying to DoDual Level of Analysis

• Adaptionist Level (Information Scent)
• Why of action
• Problem user is trying to solve
• Specify without knowing mechanisms
• Proximal Mechanism Level (Problem Space Rules)
• Mechanisms of action
• How each step proceeds

80
Physical Layout of Patches Reduces Cost of
Work
81
information scent
Tokyo
Cues that facilitate orientation, navigation,
assessment of information value
New York
San Francisco
82
spreading activation
Base-level reflects likelihood of occurrence
Strength of link spread reflects likelihood of
cooccurrance
83
The Information Big Bang