Transcript PowerPoint - Carnegie Mellon School of Computer Science

Augmenting Groupware with
Intelligence:
Supporting Informal Communication,
Trust, and Persona Management
Joe Tullio
Dissertation Proposal
May 1, 2003
2
Overview
Introduction/Motivation
Informal communication and calendars
Intersection of CSCW, IUI, Ubiquitous computing
Thesis statement/Contributions
Related work
Proposed work:
Augmenting calendars with attendance predictions
Effects of augmented calendars on user attitudes and
behaviors
Strategies for managing persona in groupware – a taxonomy
Timeline for completion
3
Calendars as tools for informal
communication
Definitions: GCS, Informal communication
Studies:
Palen, L. (1999) "Social, Individual & Technological Issues for
Groupware Calendar Systems", CHI'99.
Grudin, J. and Palen, L. (1997) "Emerging Groupware Successes in
Major Corporations: Studies of Adoption and Adaptation",
WWCA'97.
“Calendar work” +
– Locating colleagues
– Assessing availability
– Regulating privacy
4
CSCW framework (Dix 1994)
Two people and a shared
artifact
People interact with one
another and with the
artifact
People even communicate
through the artifact
5
Calendars in the CSCW
framework
Calendar as the shared
artifact
People communicate
informally
People maintain and
browse calendars
People communicate a
persona through the
calendar
IUI/User modeling
Calendars can be inaccurate
Wrong recurrence boundaries
Conflicting events
Infrequently attended events
We have a basis for modeling
Domain knowledge
Attendance history
There is uncertainty in event attendance
Inherent error & misrepresentation
Bayesian networks
7
Ubicomp
Leverage mobile devices
Individual practices
Ubiquity of calendars
Calendar itself can be used as a sensor in
context-aware applications
8
At the intersection
Calendar is an artifact supporting informal
communication
Mobile, individual calendars can exhibit
inaccuracies
Inaccuracies can be mitigated with intelligent
assistance
Calendar representation can be seen as persona
Persona – One’s representation through a shared
artifact
Intelligent systems can diminish control over
this persona
9
Proposed research
Build a prototype groupware calendar that
incorporates attendance prediction
Use this prototype to explore:
Feasibility of a Bayesian model
Effect on user communication practices
Effect on user attitudes toward adoption, trust
Develop a framework for persona management
Groupware augmented with intelligence
Focus on learning
10
Thesis statement
The GCS’s role as a tool for computer-supported
cooperative work can be better supported through the
application of predictive user models. These models can
improve it as a predictor of user activity and
consequently as a facilitator of informal communication.
I can validate this claim through an exploration of its
use in a real-world setting. I can then develop a
taxonomy of techniques for managing the persona
conveyed by such artifacts along dimensions of the
broader class of intelligent groupware applications.
11
Research contributions
Technological solution to the problem of
inaccurate calendars
Impacts many context-aware applications
Analysis of feasibility
Socio-technical effects of this solution
Identify changes in communication patterns
Examine user attitudes toward intelligent assistance
Persona management
Framework for designers and researchers
Ground intelligent groupware to the social needs of
the workplace
12
Related work:
Studying calendars
Aforementioned work by Palen, Grudin
Academic environment – Mitchell
Ubicomp systems with calendars:
Horvitz et al - Priorities
Tang et al – Awarenex
Marx et al - CLUES
13
Related work:
Intelligent groupware
These systems use different representations, UIs, and
learning algorithms…
Challenging in terms of evaluating their
Effects on communication practices
Influence on adoption
User trust, especially in early stages of learning
Horvitz et al - Coordinate
Begole et al - Rhythm Modeling
Ashbrook & Starner – GPS, Markov models
Hudson et al – Predicting availability with sensors
14
Related work:
Learning/Trust/Persona
Mechanisms to support trust
Some initiated by users, others implemented by designers
Assist in learning and to manage appearance
Plausible deniability
Can use impoverished information, system error to justify
absence
Maes – Agents for email, meeting scheduling
Tiernan/Czerwinski – Notification agents
Farnham – Social networks
De Angeli et al – Biometric verification
15
Overview
Introduction/Motivation
Informal communication and calendars
CSCW, IUI, Ubiquitous computing
Thesis statement/Contributions
Related work
Proposed work:
Augmenting calendars with attendance predictions
Effects of augmented calendars on user attitudes and
behaviors
Strategies for managing persona in groupware – a taxonomy
Timeline for completion
16
Augur: A probabilistic shared
calendar (Goecks, Nguyen)
Calendars shared from personal mobile devices
Support individual practices
Probabilistic model predicts future attendance at coscheduled events
Make the calendar a better predictor of activity for both
workgroups and context-aware applications
Visualize predictions in a browsable calendar
Awareness for informal communication
From Ambush: Support for “ambushing”
17
Representation: Bayesian
networks
Compact, descriptive representation of a
domain with uncertainty
Need domain knowledge, some structure
Capable of learning over time
Capable of generating explanations if
needed
Augur Bayesian network
Augur system
architecture
Augmented personal calendar
21
Results to date
Ambush: Stabilization on routine events
SVMs used to identify role, location, event type.
Event Type 80%
Location 82%
Role – more participants needed
Publications:
Tullio, J., Goecks, J., Mynatt, E., Nguyen, D. Augmenting
Shared Personal Calendars. UIST 2002.
Mynatt, E. and Tullio, J. Inferring Calendar Event
Attendance. IUI 2001.
22
Feasibility of Augur
Do predictions converge with actual
attendance over time?
What type(s) of events perform better?
Is the model’s structure appropriate?
Completeness
SVMs for event classification
23
User attitudes and behaviors
Study Augur’s ability to support informal communication
Study attitudes toward trust, adoption
Building on the work of corporate calendar studies at
Sun, Microsoft, Boeing and others
Also designing to the practices of our academic
environment
Ambushing
Personal (PDA-based) calendar practices
Noisy calendars
24
Evaluating attitudes/behaviors:
Proposed activities
Four deployment phases
1. Preliminary (Summer/Early Fall 2003)
•
•
Initial attitudes and practices
Interviews
2. Calendar deployment (Early Fall 2003)
•
•
Collecting training data
Let users become accustomed
3. Intelligent calendar deployment (Late Fall 2003)
•
•
Investigate changes in attitudes/practices over time
Collect measures in accuracy
4. Persona management (Spring 2004)
25
Participants
Study group
20 participants from several FCE labs
Both “readers” and “writers”
Some working closely, others infrequently
Periodic interviews before, during, after deployment
Larger pool of readers
Expecting advisees, students in courses to read
Log accesses to identify browsing patterns
Limit reading to school machines
26
Challenges
No existing GCS infrastructure
Ramp up by first using shared calendar without
predictive features
Must design for possibly several common tools
Dynamic schedules and personnel
Some learned patterns are incompatible with
changes in term/personnel
Attitude changes versus behavior changes
Opinions may change without measurable changes
in activity
27
Why FCE is promising
Open environment
Not subject to closed calendars at higher positions in
the hierarchy
Existing calendar habits
Personal, “noisy” calendars the norm
Individual calendars demonstrate need for intelligent
assistance
Abundance of events/activities
Ambushing
Seems to be a common practice
28
Method - Interviews
Augur as a communication tool (behavior):
How often did you check your calendar/others’
calendars? To what purpose, if any?
How often did you use the predictive features?
Did you change your calendaring habits?
Trust in intelligent assistance (attitude):
How accurate were the predictions for others?
Did they seem to improve or degrade?
Were you represented accurately?
Did you attempt to change your own predictions?
29
Observing behavior
Log calendar accesses
Browsing up/down the hierarchy
Confidentiality
Semi-controlled situations
Present interviewees with task using the
calendar, see how they would accomplish it
Control for same calendar information
Think-aloud
30
Success metrics
System sees use
Neglect over the first weeks may necessitate
new incentives (or setting)
Changes in behavior observed through
logs and interviews
Changes in attitude evidenced through
interviews
31
Persona management
Persona – One’s representation
through a shared artifact
Management implies
negotiation with the system
Common property of groupware
in general
Complicated by intelligence
Important in early stages of
learning
32
Management strategies
Strategy
Social norms
Preferences
Gaming
Learning
Editing of output
Description
Users develop new social practices to
deal with a shared artifact.
Users configure settings that control
how their informat ion is gathered and
used.
Trial-and-error manipulation of inputs
eventually results in learning a
working conceptual model of the
system.
Preferences are learned over time by
unsupervised or example -based
techniques.
The system’s output is edited directly
by the user to a desired state.
Changes are possibly fed back into
the system’s user model.
Example
Plausible deniab ility of instant
messenger response.
Access-control lists, privacy
settings in shared calendars.
Using particular key words in
email to boost its priority.
Automatic scheduling systems
for shared calendars.
Customized status
informat ion on instant
messaging clients.
33
User dimensions
Dimension
Privacy preferences
Collocatedness
Frequency of interaction
Description
User’s attitude toward what
information should be shared and what
should be kept private.
To what degree a user is geographically
near her colleagues.
To what degree a user interacts with
her colleagues in the workplace.
Effects
May impose restrictions on use of
personal information and thus
permit a less rich persona.
Collocated users may rely more on
social norms to resolve issues
concerning persona.
Coworkers with little interaction
may rely more on technological
means of persona management.
34
Application dimensions
Dimension
Number of inputs
Description
The variety and amount of personal
information used by the system.
Number of outputs
The number of co mponents comprising
the machine-generated persona.
Accessibility of inputs
Ease with wh ich users can control the
information used by the system.
Complexity of algorithm
Co mplexity of the process which
transforms personal information into a
shared persona.
Example
Many:
Fogarty et al [32]
Few:
Instant messenger
status
Many:
Augur
Few:
Instant messenger
status
Accessible: Calendar events
Less
A/V sensors
Accessible:
Simp le:
Basic open-access
calendar
Co mplex:
Augur
35
Method
Populate framework with existing systems
In particular, look for examples of
complex, intelligent groupware
Augur system: support persona
management using framework
Deploy after term change to explore use
Mitigate misrepresentation from error
36
Success metrics
Dimensions of framework
Does the body of intelligent groupware divide this
way?
Is the problem much more complex?
Utility
Does the taxonomy identify new research areas?
Does it provide design guidance?
Are Augur users able to successfully manage their
representation through the shared calendar?
Interviews from previous phase
37
Timeline
Fall/Spring 2000:
Development/Test of Ambush system (published at IUI 2001)
Fall/Spring 2001:
Development/Test of Augur system (published at UIST 2002)
Summer 2003:
Prepare Augur for deployment, develop interview questions, solicit participants
Literature review/development of framework
Fall 2003:
Deployment/Evaluation of Augur
Begin design/implementation of persona management for Augur
Spring 2004:
Re-deploy Augur with persona management features
Take advantage of schedule change
Summer 2004:
Writing and defense
Acknowledgments
Thanks to Elaine Huang, Jeremy Goecks, David Nguyen,
my committee, the Everyday Computing Lab, and
everyone else who discussed or critiqued this work with
me.
Thanks also to the National Science Foundation
CAREER Award #0092971.