Personalized Information Services

1
Personalized Information Services

• Javed Mostafa
• Indiana University, Bloomington

2
Outline

• Personalization as part of a broader field
• Personalization vs. customization
• Representation
• A research issue in personalization
• Approaches taken to study the issue
• Results
• Conclusion

Acknowledgment The research described in this
presentation is a collaboration among a number of
people. I am grateful for work conducted by Dr.
Mukhopadhyay Dr. Palakal (Computer
Information Science, IUPUI). I am also indebted
to two of my previous students Luz Quiroga and
Junliang Zhang. Thanks also to NSF for
funding this research.
3
Connection to a broader field

• Personalization is part of a larger field known
  as context aware computing (CAC)
• CAC is concerned with a broad range of problems
  including development of smart environments
  (offices, homes, cars, etc.), smart weapons and
  appliances, smart clothing, and information
  systems
• Some interesting projects
• Project Oxygen (MIT) http//oxygen.lcs.mit.edu/Ov
  erview.html
• SmartSpaces (NIST) http//www.nist.gov/smartspace
  /smartSpaces/
• Adaptive Systems Attentive User Interfaces
• (Microsoft) http//www.research.microsoft.com/ada
  pt/

4
Context Aware Information Services (CAIS)

• Goal Basic information support services (i.e.,
  browse, search, filter, presentation and
  visualization) should be seamlessly available
  from any location, any device, or any
  application, and in a form that permits optimum
  use of the information

5
Context Aware Information Services (CAIS)

• Context is complex
• Users can interact with
• a variety of info systems their desktop, a
  laptop, a handheld, or a palmtop
• A variety of applications and documents
• Users may be stationary or mobile

6
Levels in CAIS
Tablet
Desktop
PDA
MS-Word
Photoshop
MS-Excel
Netscape
Different types of documents and content
Users interaction, users short term demands,
user s long term needs
7
Requirements Proactive awareness and responses

• Proactively seek information related to content
  being manipulated by the user and bring related
  and relevant information to the users attention
• Automatically modulate the features and
  presentation according to device and application
  characteristics

8
Contexts of a Typical User
Location
Device
Applications
Tasks
Information
Immediate and long-term info demands
9
Customization vs. Personalization

• Customization taking into account contexts
  other than those that represent personal
  information demands and interests (short- or
  long- term)
• Personalization taking into account contextual
  information related to users information demands
  and interests (e.g., query terms, relevance
  feedback on documents, rating, etc.)
• Both, together, support context aware information
  services

10
Customization vs. Personalization
Location
Representation for customization
Device
Applications
Tasks
Representation for personalization
Information
Immediate and long-term info demands
11
Representation

• To provide context aware info services requires
  maintaining up-to-date contextual information in
  a form that permits efficient computation and
  accurate predictions about users info needs,
  i.e., need context representation

12
Representation for Personalization User Profile

• We developed a representation to predict
  relevance of new information according to users
  interest and long-term information need
• Requirements supported
• Online learning
• Low latency
• Permits exploration and adaptation

13
Generating the representation

• To generate the representation we relied on
  rating or indicators of interest on topical
  categories
• The representation contained two types of
  information topical categories and assessment of
  interest in the categories

14
Interest representation for personalization
Probability that category 2 is the most relevant
category
Probability that category 1 is relevant to the
user
u1
t1
u2
t2
u3
t3


un
tn
Documents
Top class
Relevance of categories
User profile/model
15
Source of interest information

• Explicit Users were asked to provide rating on
  documents
• Implicit Users interaction with content and the
  interface were taken into consideration
• Such interest information was converted into the
  (two-level) profile/model by using a simple RL
  algorithm
• Mostafa et al. A multilevel approach to
  intelligent information filtering Model, system,
  and evaluation. ACM TOIS, 15(4), 1997.
• Different applications have been created, incl.
  SIMSIFTER and TuneSIFTER
• See lair.indiana.edu/research/

16
Research issue Big picture

• Interested in two types of research issues
• With any type of intelligent HCI a fundamental
  issue is control
• Who is in charge?
• If the user wishes to delegate, how much autonomy
  should the system have?
• Agent vs. User (Direct Manipulation)
• Maes Shneiderman debate http//www.acm.org/sigc
  hi/chi97/proceedings/panel/jrm.htm
• If the user wishes to take charge, how much
  responsibility should the user take on? User
  effort user involvement can impact system
  effectiveness

17
A research issue Users Role in Personalization

• Type of interest
• Interest change
• User Involvement
• Amount of interaction
• Type of interaction

18
Approaches to study the research issue

• As it is v. difficult to manipulate certain
  conditions (e.g., change of interest w.r.t.
  certain topics) we developed a simulation tool
• For other conditions we conducted experimental
  studies with actual users

19
Simulation study using SIMSIFTER

• Type of interest may impact the rating (degree
  and frequency)
• Rating may impact how quickly the system can
  learn or generate an accurate profile
• Accuracy of profile determines accuracy of
  prediction of relevance
• SIMSIFTER used about 1.4K consumer health
  documents and 15 categories of health information
  (anxiety, allergy, heart, cholesterol,
  depression, diet, environment, exercise, eye,
  headache, lung, medicine, teeth, men-health, and
  women-health )

20
Study Different Profile Types

• We created different types of profiles
  concrete, middle, and mild-low
• Degree of interest was used to generate rating
  probabilistically
• Frequency of rating increases with increased
  intensity of interest

21
Results Different Profile Types
Impact of different types of interest on
prediction of relevance
22
Study Change in Interest

• Over time as the user is exposed to continuous
  flow of new information and users situation
  changes, the user may experience change in
  interest
• Change in interest may be gradual or abrupt

23
Results Change in Interest
Impact of change in interest on prediction of
relevance
24
Study Modalities of interest information
collection

• Interest information can be collected explicitly
  by asking the user
• By generating the rating based on content viewed
  by the user
• Or, a combination of both of the above strategies

25
Results Different modalities of interest
information collection
Impact of different interest information
collection modalities on prediction of relevance
26
TuneSIFTER Study

• Aim was to engage actual users and analyze
  different modalities of interest information
  collection
• Rule-based
• Explicitly by requiring users to rate
• Implicitly by observing behavior and associating
  behavior with rating
• Provided access to music titles in a dozen genre
  from the MP3.com service
• 35 subjects recruited from IUB

27
TuneSIFTER User Interface
28
Study Three modalities of interest information
collection

• Rule based user provided the profile in the
  first session
• Explicit learning user rated music titles
• Implicit learning different sources used
  users click on the column of title, users click
  on the column of artist name, users click on the
  column of genre, and users click on the column
  to request more information. In addition, the
  time user spent on listening to the music was
  also recorded by the implicit-learning system

29
Results Three modalities of interest information
collection
30
Conclusions

• The representation and the learning approach
  developed are quite robust in terms of capturing
  different types of interest and change in
  interest
• Implicit modality, when time data is available,
  may be applicable in reducing user involvement
  without sacrificing performance

31
Limitations and Future Work

• User involvement may vary with tasks and domains
• For example Kelly and Belkin (2002) state that
  reading time is not a reliable source for
  implicit modeling
• Different levels of modeling may be needed
• Topical granularity in the user profile
  influences performance Quiroga and Mostafa
  (2002)
• Two-level modeling needed in the News domain
  (content highly dynamic)

32
Additional Citations

• Kelly and Belkin. Modeling characteristics of the
  Users Problematic Situation with Information
  Search and Use Behaviors. JCDL Workshop on
  Document Search Interface Design,
  http//xtasy.slis.indiana.edu/jcdlui/uiws.html,
  2002.
• Quiroga and Mostafa. An Experiment in Building
  Profiles in Information Filtering The Role of
  Context of User Relevance Feedback. Information
  Processing Management, 38(5), 2002.
• Pitkow et al. Personalized Search. CACM, 45(9),
  2002.
• User modeling 10th Anniversary Issue. Gerhard
  Fischers work in this area is especially
  recommended.

33
Related IR Forums

• SIGIR - ACM Special Interest Group on Information
  Retrieval Conference
• UIST  - ACM User Interface Software Technology
  Conference
• UIU - ACM Intelligent User Interfaces Conference
• TREC - Text REtrieval Conference
• ASIST - American Society for Information Science
  and Technology Conference
• JCDL - Joint Conference on Digital Libraries
• CIKM - Conference on Information and Knowledge
  Management
• AGENTS - International Conference on Autonomous
  Agents

34
Need more information?

• Our lab
• Laboratory of Applied Informatics Research
  (lair.indiana.edu)
• Email jm_at_indiana.edu