Information Filtering User Modeling Lecture 05

1
Information Filtering User Modeling Lecture 05

• Tsvi Kuflik
• Department of Information Systems Engineering
  Ben-Gurion University of the Negev
• Beer-Sheva 84105, Israel
• tsvikak_at_bgumail.bgu.ac.il
• Based partially on Users are individuals
  individualizing user models, by Elaine Rich,
• International Journal of Man-Machine Studies
  (1983) 18, 199-214

2
IF Model
Goals, tasks Long term User with
Producers of Texts
Information Interests Regular
Distributers of Texts
Representation and distribution
Representation
Profiles
Text Surrogate
Comparison or Filtering

• Retrieved texts

Use and / or evaluation
modification
3
User Modeling

• In order to allow comparison of Information
  (data-items / documents) to user interests we
  need
• Data-Items representation - traditional IR
• User information needs definition - traditional
  IR and beyond - AI

4
User Model within IF Model
Goals, tasks Long term User with
Producers of Texts
Information Interests Regular
Distributers of Texts
Representation and distribution
Representation
Profiles
Text Surrogate
Comparison or Filtering

• Retrieved texts

Use and / or evaluation
modification
5
User Modeling

• What is a user model?
• The general area is information filtering
• Practical definition
• the knowledge and inference mechanism which
  differentiates the interaction across
  individuals (James Allen)

6
User Modeling

• Scenarios where user model may be relevant
• When an individual is not familiar with the
  details needed to perform a task
• To save time (avoid looking through all search
  results)

7
User Modeling

• User model emphasizes information about the
  person
• The task performed and the context it is
  performed in have significant impact as well
• What about demographic details?

8
User Modeling

• Elaine Rich User Model definition (work done late
  70s early 80s)
• There are 3 dimensions to user model space
• Canonical user vs individual models
• Explicit vs Implicit definition
• Long term vs Short term characteristics

9
User Modeling

• Canonical user vs individual models
• Generic user model
• Predefined
• No personal data
• Stereotypic user model
• Some personal information
• Selection of the most suitable model
• Individual
• User personal data
• All may be adaptable

10
User Modeling

• Explicit vs Implicit definition
• Explicit definition
• Accurate information
• Tedious
• Implicit definition
• Questionable accuracy
• Heuristic
• Depends upon users actions
• Both can serve as initial source that is later
  on adapted

11
User Modeling

• Long term vs Short term characteristics
• Long term
• Long lasting
• Worth user effort
• Short term
• No effort
• Generic
• Stereotype
• Implicit

12
User Modeling

• Information Filtering User Modeling
• According to Rich terminology
• Individual
• Explicit Implicit
• Long term (mainly)

13
User Modeling

• How can we model user information needs?
• How does a user profile look like?
• How can we generate one?
• How can we adjust user profile over time?

14
User Modeling

• Traditional IR represents user information needs
  and Data Items similarly
• This representation allows easy and efficient
  comparison

15
User Modeling

• Content Based User Profile
• Vector Space Model
• Document representation as vector of terms
• Weighting scheme
• Similarity

16
User Modeling

• The Rocchio classifier
• Relevance feedback Vector Space model
• ROCCHIO algorithm generates a profile that is the
  centroid of the training examples
• b, g may be set to 1 0, 1 1, 16 4 or
• wkj is the weight term tk has in document dj

17
User Modeling

• The Rocchio classifier
• Ongoing adaptation

18
User Modeling

• Content Based User Profile
• Probabilistic approach
• Profile representation
• Document representation
• Similarity

19
User Modeling

• Naïve Bayes
• Recall the term weighting
• How do we estimate initial values?

20
User Modeling

• Naïve Bayes
• Estimating qik
• Assume the number of irrelevant documents in the
  collection collection size
• N collection size
• ni - number of documents with term ti
• qik

21
User Modeling

• Naïve Bayes
• Estimating pik
• Assume global value p for all pik
• Term weighting by IDF
• ?BIR(qk,dk)
• Often used p0.5?cp0

22
User Modeling

• Naïve Bayes
• Adaptation using relevance feedback
• User views documents and respond with relevant or
  non-relevant decision
• r number of documents judged as relevant for qk
• ri number of documents with term ti
• Improved estimation

23
User Modeling

• Information Filtering User Modeling
• Time span
• Long for getting a trained system to work
• sometimes medium and even short term interests
  exist
• easier with user similarity then content

24
User Modeling

• Information Filtering User Modeling
• Individual system
• Tailor made
• All relevant aspect taken into account, including
  
• area of interest
• background (socio-demographic) information
• personal preferences
• similar users
• more

25
User Modeling

• Information Filtering User Modeling
• Rich terminology
• Individual
• Explicit Implicit
• Long term (mainly)
• Other aspects
• Socio-demographic
• Rules of thumb
• Collaboration

26
User Modeling

• Content Based User Profile
• Problems
• Over specialization

27
User Modeling

• Collaborative User Profile
• Basic idea
• Similarity of responses
• Implementation
• Collect user responses to various data items
  presented over time
• Generate a preference matrix of users and items
• Use similarity of users in order to recommend new
  items to users

28
User Modeling

• Collaborative User Profile
• Problems
• First reader problem
• The first reader of an item will not benefit from
  rating it, so he/she will not respond
• Response matrix sparseness
• Many items
• many users
• As a result sparse matrix

29
User Modeling

• Collaborative User Profile
• Summary
• Simple idea
• Logical idea
• Works (?)

30
User Modeling

• Rule based User Profile
• Basic idea
• Knowledge representation
• Rooted in early AI systems
• Easy for implementation
• Captures every relevant aspect of user knowledge
• Content and meta data
• Personal preferences
• Easy for implementation

31
User Modeling

• Rule based User Profile
• Problems
• Inaccuracy of user definition
• Information needs definition is problematic
• Effort required (definition and maintenance)
• Users dont like to put a lot of effort in
  training systems

32
User Modeling

• Traditional IR represents user information needs
  and Data Items similarly
• This representation allows easy and efficient
  comparison
• But
• Users are more then list of keywords
• Many User-aspects can not be represented as list
  of words

33
User Modeling

• What are todays user information needs?
• Professional
• Knowledge acquisition
• Recreational
• Shopping
• General Information
• Anything else...

34
User Modeling

• What are todays tools for information gathering?
• Internet
• Search engines
• Directories
• Meta searchers
• All the above requires some (or a lot of) effort
  and provide some (or a lot of) junk

35
User Modeling

• What is needed?
• Better personalization
• Systems should know user interests
• Systems should react accordingly -provide
  relevant and only relevant information

36
User Modeling

• How can it be achieved?
• Teaching / training the system for initial
  activation
• Ongoing modification, based on performance

37
User Modeling

• Information Filtering User Modeling
• Implicit definition
• Look over the users shoulder
• Constant adaptation
• Hard to start
• Explicit definition
• Tedious
• Using example
• Possible / optional explicit definition

38
User Modeling

• Application areas
• Any personalized service
• Name some
• Any personal device
• Name some

39
User Modeling

• Application areas
• What can be personalized
• Content
• Interface
• Visualization
• Activation

40
User Modeling

• Application areas
• How it can be personalized
• Intelligent interface agent
• Works in parallel with the user
• Monitor users behavior
• Infers, builds and updates user model
• Filter incoming information