The Doppelgänger User Modeling System

by Jonathan L. Orawant

Slides

• Continuous ---> Frequent / Infrequent

• Unintrusive ---> Passive / Active

• Modularity

• Sensor confidence values

• Badger

• Newspace

• UNIX aliases

• UNIX lastcomm

• UNIX finger

• UNIX calendar

• free text

• Doppelganger

• ... other clients

• SPONGE knowledge representation language

• Assertion confidence values

• Inducer fills in gaps with weighted averages from the current database ---> Communities [not Stereotypes]:

• membership is not binary

• defined by constituents

• user can be in many communities

• Trust
  • explanation is the key to earning trust

  • simple agents trusted over intelligent agents

• Conversational mode
  • the system is proactive, so it has something to say

  • "all users can be reached via electronic mail"

  • natural language interface

  • annoyance parameter

• Editor mode
  • graphical user interface for non-technologists

  • colors, lengths, and thicknesses

• Behavior of clients

• Application may reveal information from the user model without explicitly citing the model
  • e.g. newspaper gives one line reason for including a particular article

• Markov models
  • State diagrams where arrow thickness indicates probability of each transition

• Linear prediction
  • Bar graphs indicating frequency and duration (e.g. predicting when the user will be logged on)

  • Future behavior predicted from most recent data, confidence in predictions indicated by saturation

• User preferences in Newspace
  • Sliders change from gray to bright red as the value deviates from the population norm

  • Highlights the model's "most interesting traits" Adapting the Interface to the User Model

  • "The ideal interface to a user modeling system would itself make use of a user model to gauge the efficacy of its different techniques."

  • "... novice users would view the pictures shown in this paper, and hackers might edit the LISP-like models themselves."

Questions for Discussion

1. Do we agree with Orwant that there is a trade-off between more intelligent (and therefore speculative) inferences and the user's trust in the system? (pg.151)

2. In the Newspace preference interface, how useful is it to highlight the traits that differ from the norm? (pg.155)

3. Would Doppelganger benefit from an explicit knowledge base? Where would such a base come from? How would this affect the user's trust? (pg.5)

4. Would you trust a system like Doppelganger?

Class Discussion

Sensors

• some of them not useful? How to tell if person is acutally using the sytem? (ie. how to tell if 5 minutes spent on a news article meant the user was reading it and interested or they were interrupted by a call or some other person and just skimmed the article - thus implying the user is not interested...)

• difficult to find out the internal details of the system
  
• comparison to Open Sesame

• collecting some information not worth what you get out of it
  
• seems to be just hooking up sensors because he can (use of the badge a good example?)

• isn't it scary, intrusive (that is, someone being able to access such info about you)?
  
• slam on MIT students - probably hacking lax security :-)
  
• point: a lot of info can be inferred from simple models, but others are hard (newsrc vs. actual news read analysis).

• slam on news reading systems telling exactly what was read - there's more than 1 article on the page (which did the user read and pay attention to?)

• student questions trust issue:
  • trust system's modeling (covered by Orwant)
  • trust what system does with model
  
  
• prof says not part of Orwatt's area. It's up to the individual application to decide what to do with the model
  • questions the security of private data -> system makes use of others' models (although you can't change others' models).

Intelligent Agents

• prof: important question -> "Is an intelligent agent less trustworthy than a dumb one?"

• student: As long as I know what/how it functions.
  
• prof: [and] trust the information that it gives back
  
• student: could be a query problem, student-modeling problem
  
• prof: inference base could be wrong - not trustworthy.
  
• student: wants agent to justify its inferences about the user and control to adjust/stop bad inferences.
  
• student: take care with form of query
  
• student: know/be aware the agent is guessing

• prof: "does everyone know they're being modeled?"
  
• student: it sends email and you have to turn it on

Question 2

• student: if you take the norm, you may get unwanted info (OJ example)
  
• student: info from norm is useful - at least to be aware of the norm
  
• prof: would you like to have a system classify you as strange?

• student: like to see system's feedback about news choices
  
• student: system's classification could affect user action
  
• prof: do people try to fit a model (allusion to Coca-Cola advertising and its success) ?
  
• student: Ad companies manipulate such beliefs/feelings. What if system lied about the norm?
  
• prof: comment on "treacherous" systems that use user models for ill gains
  
• student: e.g. the Press's concern over net-specialization - soln: net-randomizer

Question 4

• prof: would you trust the system?
  
• student: let's talk about Question #3, communities (extended communities like seen in the GUMS model)

• prof: examine inference engine in Doppelganger -> assumption of unknowns based on likeness. Ok., works in real life, why not in a system?
  
• student: brought up point of using other's models (viewing world though other's paper -> "the all OJ paper" :-)

• prof: back to the question, would you use a system where your UM is not private?
  
• student: questions use of the name Doppelganger (threatening, and Adam's note: too hard to type that ä character).
  
• student: [good] point, not all [of it], I'd trust it to get info on similar users
  
• prof: yes, but I'd have to build up trust in such a system

end of discussion