PROBLEM BEING ATTEMPTED

1
PROBLEM BEING ATTEMPTED

• Privacy-Enhancing Personalized Web Search
• Based on
• User's Existing Private Data
• Browsing History
• E-Mails
• Recent Documents
• All Data Stored on User's Personal Computer
• Classify Interests into 2 Categories
• General Interests (Less Sensitive to Privacy)
• Specific Interests (More Sensitive to Privacy)

2
PROPOSED SOLUTION

• Automatically build user profile from available
  source data
• Similar Terms Two terms that cover the document
  set with heavy overlaps might indicate the same
  interest area
• Use Jaccard Similarity
• Parent-Child Terms Specific terms often
  appeartogether with general terms, but opposite
  not true.
• e.g. Badminton and Sports
• Use Conditional Probabilities
• Control the information sent to the search engine
• minDetail Determines which part of the user
  profile is protected
• expRatio Measures how much private information
  is exposed to the server
• Inversely Related
• Wrapper to personalize results
• Use previously constructed profile
• Parameter a that determines amount of
  personalization
• PPRank a PersonalRank (1-a) SearchEngineRank

3
CRITICISMS

• Number of Returned Results Considered
• Only top 50 results from the original search
  engine results considered, but user's preferences
  could project a lower result into the top 10 ?
• Measure of Search Quality
• Average Precision is used but authors fail to
  explain how this ties in with the quality of the
  personalization in the search
• Testing the Effect of Manual Privacy Settings
• The authors don't provide experiments that test
  how the system would behave if users, instead of
  using minDetail all the time, manually excluded
  some terms from their profile
• Classification of Terms
• A specific term might be classified wrongly as
  general if it occurs often enough in the user's
  corpus but it is extremely revealing as far as
  privacy is concerned

4
RELATIONS TO COURSE TOPICS

• Personalized Search
• Query Disambiguation The authors use a variant
  of the course example searching for Rockets, a
  sports fan wants the basketball team, not links
  related to space exploration
• Relevance Feedback This system is implicit in
  seeking relevance feedback through the user's
  browsing history and other personal information,
  rather than explicit.
• Dealing with Unstructured Data
• Browsing history, e-mails and recent documents
  all unstructured
• New system to classify terms into either
  similarity or parent-child relationships based on
  occurrences
• Term Similarity
• Used to classify terms use Jaccard Similarity
  and Conditional Probability
• Weighted Re-Ranking of Results
• Similar idea, but not same, as PageRank
• Get the results from search engine for given
  query
• Then re-rank the results based on user's personal
  profile, using weighting factor a

5
(No Transcript)