User Modeling via Stereotypes

1
User Modeling via Stereotypes

• Elaine Rich
• University of Texas

2
Why User Modeling?

• People need to form a model of the person with
  whom they are dealing before they can behave
  properly.
• Examples
• PS2 Controller size, Auto Insurance Quote

3
User Modeling via Stereotypes

• A stereotype is a cluster of characteristics
• e.g.
• Profession Stereotype
• income level, dress code at work, degree, etc
• Ethnic group stereotype
• Food preference, spacial distance

4
Stereotypes Are Based on Domain

• Entertainment Industry
• caremovie, music preference, education level
• dont dress, weight, height
• Car Sales Industry
• care income, family size
• dont education level, religion

5
Stereotypes Are Based on Probability

• How confident are we ?
• Computer Science Professor - NON_TV 85
• College Graduates - over 20 90
• Probability may result in unfairness.

6
So, a Stereotype is a set of triples

• (Attribute, Value, Rating)
• Medical Doctor
• (Income, 4, 900)
• (Education, 5, 900)
• (Afraid-of-Blood, -5, 900)
• (Watch-TV, -3, 800)

7
A Stereotype Contains Triggers

• A trigger is a hint that instantiate other
  stereotypes.
• Medical Doctor
• High_Education_Trigger -gt NonTV_StereoType
• High_Income_Trigger -gt House_Owner
• A trigger also has rating as confidence level.

8
Over View of Grundy the Librarian

• Has pre-built stereotypes
• Hierarchical memory
• Global, individual, dialogue
• Algorithm for activating stereotypes
• Adaptation of stereotypes

9
Activating Stereotypes (by triggers)

• If trigger already instantiated, do nothing, if
  not, instantiate it.
• Example
• Name John - Man_trigger (instantiate)
• Father - Man_trigger (ignore)

10
Activating Stereotypes (by triggers)

• If stereotype has not been activated before, it
  is activated now.
• Name John -gt Man_trigger -gt Man_stereotype
• If stereotype has been activated before and still
  active (confirmation)
• Father -gt Man_trigger_Rating up -gt
  Man_Stereotype_Rating up

11
Activating Stereotypes (by triggers)

• If stereotype has been activated before, but its
  not activated now. The situation must be
  re-examined on the basis of the balance of the
  evidence is in favor or opposed to the
  stereotype.
• How to calculate the balance of evidence???
• Whats the definition of in favor or oppose
  ?

12
Adaptation of stereotypes

• Why?
• Lack of real data at construction time results
  in the need of adaptation of stereotypes.
• Concern
• Prevent bias because of frequent usage by a few
  users, need weighted constant of the overall
  values.

13
Adaptation of stereotypes

• Confirmation Case
• increase VALUE
• increase RATING
• M.D. -gt (Income, 3, 800)
• Drive Ferrari -gt (Income, 4, 900)
• New_Value (3 100 4) / (100 1) 3.01 gt 3
• New_Rating 800 900/800 801.125 gt 800

14
Adaptation of Stereotypes

• Conflict case
• decrease VALUE
• decrease RATING
• M.D. -gt (Non_TV_Person, 4, 800)
• Subscribe to TV_Guide -gt (Non_TV_Person, -5,
  900)
• New_Value (4 100 - 5) / (100 1) 3.91 lt 4
• New_Rating 800 - 900/800 799.875 lt 800

15
Results of Grundy
Controlled Random
GOOD 102 54
BAD 39 60
16
The End