Transcript Analysis - Courses

The Cost-of-Knowledge
Characteristic Function:
Display Evaluation for Direct-Walk
Dynamic Information Visualizations
CHI ‘94
Card, Pirolli, & Mackinlay
Concepts



Information has cost structure
Objective: maximize information benefits
per unit cost (cost ~ time)
Cost-of-Knowledge Characteristic Function:

Characterizes the effect of a design of dynamic
display/human-computer dialogue on
information’s cost structure
Cost-of-Knowledge
Characteristic Function

Improve productivity: Less time or more
output
Case study

Direct-walk interactive infoviz




Navigate an information structure using mouse
points/other direct manipulation methods
Analyze 2 calendar programs: Spiral
Calendar vs. Sun’s CM
Users: 4 users for each study, 2 overlapping
Task: navigate to another day in calendar
Steps To Construct Cost-ofKnowledge Function


Use tasks that take different amount of time to
obtain different amount of information
Identify cost drivers for the tasks:





In Spiral Calendar: # of cycles to go through
In CM: # of different steps
Measure time taken to perform each task as cost
Perform regression of time (cost) as dependent
variable and cost drivers as independent variables
Plot cost vs. amount of information that can be
obtained
Cost Drivers

Spiral Calendar: Number of display cycles
(Century, Decade, Year, Month, Week, Day)
selected


Regression fn : Time = 3.3 + 3.5 * Ncycles
CM:
 m: point, menu pull-down, & select
 P: point & select
 B: press a button

Regression fn: Time = 1.3 + 3.9 m + 1.4 P + 0.36 B
Spiral Calendar Result
Computation
Sun CM Result Computation
Cost-of-Knowledge Functions
Value of Tasks

Values of tasks:





Frequency
Importance
Etc.
Needs to weight tasks by their values
Ex. Use probability density function to
weight tasks by frequency of use:
Pr{needed|D days ago}=0.34/(0.34+D0.83)
Expected Probability-Weighted
Costs
Summary



More measurable/computable method to evaluate a
design
Know your priority/objective: sometimes perceived
speed is more important than actual speed
Issues:
 How to accurately identify and measure all cost
drivers of a task, e.g. # of items?
 What if there are more than one way to perform a
specific task?
The WebBook & Web Forager:
An Information Workspace for
the World-Wide Web
CHI ’96
Card, Robertson, and York
WebBook & Web Forager
 Two
related designs
 WebBook - 3D interactive book of
HTML pages
 Web Forager – an application that
puts WebBook and other objects in
a 3D hierarchical workspace
Based On…



Cost structure of information workspaces –
the web has a uniform cost structure
Information foraging theory – users often
seek strategies to increase the encounter
rates of relevant information
Locality of reference – users tend to interact
repeatedly with small clusters of
information, and therefore keeping the cost
of accessing low
Problems At That Time…



Hotlist – still have to wait for slow access
times, not tunable to a reasonably coststructured workspace.
Multiple windows slow users down since they
overlap.
Users can only be at one page while the way
the users actually work with information is
to have multiple pages simultaneously
available at hand.
WebBook

Use book metaphor (animated 3D): next &
previous links analogous to books, familiar,
effective display

Any collection of preload pages
Can be bookmarked, put on a shelf
Various way to collect URLs: relative-URL,
Topic, Hot List, Search Reports


Web Forager





Explore the potential for rapid interaction
with large number of pages
Use gestures to increase speed with which
objects can be moved around
Focus on the web
Use a structured model to design (CoKC Fn)
3 levels: book/page  air & desk 
bookcase
Web Forager
Cost of Knowledge Characteristic
Function for Web Forager
Comments



Metaphor – do they really take advantage of
the affordances of a physical book and
workspace? What might you lose from using
this metaphor?
Speed of retrieving a web page is becoming
less an issue
Current browsers might already be able to
solve the problems posed by the authors
(and even work better, perhaps!)
Effective View Navigation
CHI ’97
Furnas
Effective View Navigation
Context
Navigate an information structure by selecting
something in the current view of the structure
Problems
 Large structures
 Limited resources of space & time
Proposed Requirements
Effective View Navigation (EVN) =
Effective View Traversibility (EVT)
+ View Navigability (VN)
Terms




View traversal: iterative process of viewing,
selecting, & moving to it
View navigation: decide where to go next
Logical graph: logical structure of the
information
Viewing graph: contains nodes that users see
in current view
EVT Requirements for Viewing
Graphs
EVT1: Small Views – space constraint
# of outgoing links of any node relative to
structure’s size must be small  small Maximal
Out-Degree (MOD)
EVT2: Short paths – time constraint
the longest connecting path relative to structure’s
size must be small  small Diameter (DIA)
EVT(G)=(MOD(G),DIA(G))
G= viewing graph
A Scrolling List
EVT=(O(1), O(n))
A Balanced Tree
EVT=(O(1), O(log n))
Improving EVT of a List

More dimensions - multi-column list: EVT=(O(1), O(sqrt(n)))
Improving EVT of a List

Fisheye sampling: EVT=(O(log n), O(log n)) –
allow jumping further, but larger view
Improving EVT of a List

Adding a tree: EVT=(O(1), O(log n)) – create categorization?
EVT Summary
Present information in a
representation that naturally supports
EVT - tree
 To fix non-EVT logical structures:



Add long-distance links
Glue with another complete EVT structure
View Navigability (VN)

Ability to find good paths to targets without
error & history-less
Terms





Outlink-info: info associated with outlink of a
node (enumeration or labeling)
To-set: all possible targets a link actually leads
to
Inferred-to-set: targets that the outlink-info
seem to indicate
Residue/scent: remote indication of a
node/target
Well-matched outlink info: inferred-to-set
implies to-set
Illustration
Strong Navigability
Requirement
1.
2.
3.
4.
Outlink-info must be
everywhere well-matched
Every node must have good
residue at every other node
Outlink-info must be small,
but need to describe the
whole to-set, not just the
next node (e.g. highway
signs)
Semantic labeling that
mirrors actual partition of tosets
Targets share residue
Good Example

Systematic labeling of trees of hierarchy,
i.e. biological taxonomy
Non-Navigable Structure



Completely unrelated/unstructured items
Only works with enumeration
Locally-related structure – no good residue
for things far away, e.g. WWW

Combine query & navigation
Combining EVT + VN

Large scale semantics (structure with larger
groups) work!




Assume n nodes & v links in the structure
Small view & diameter: v should be small
compared to n
Average size of to-sets (n/v) should be large
Carve up the semantics of the domain
efficiently – due to Small Diameter req.


Small # of intersections
Balanced hierarchy
Summary

Effective view navigation:




Small views
Reasonable # of steps to move around
Discoverable route to any target
Do navigability requirements guarantee
users to always find shortest paths?