Personalized Search on the Client Side

1
Personalized Search on the Client Side

• Shady Elbassuoni
• Prof. Gerhard Weikum Julia Luxenburger

2
About me

• Name Shady Elbassuoni
• Nationality Egyptian
• Master student in IMPRS
• Interests Computer Science, Programming

3
About my friend

• Name Lina
• Nationality Egyptian
• Geography student
• Interests Traveling, Asian Cultures

4
Query Java Map
5
Search Personalization

• Standard IR

Query
Document
User
6
Proposed Approach

• Client side web search agent using implicit
  relevance feedback
• Why Implicit feedback?
• Users do not want to do extra work
• Infer user intentions based on observable
  behavior
• Why client side?
• The users every move can be tracked
• Data completely resides on the users machine
• No extra overhead on server

7
Proposed Approach

Search Engine (Google)
user
query
Query Modification
History Log
page visits
User Modeling
query click stream
results
Result Reranking
8
Proposed Approach

• Define a session
• A set of users activities (Browsing and Queries)
  within a certain time
• A set of users activities about the same topic

9
Proposed Approach

• Combine two approaches
• If new session ? Local Index Personalization
• Else ? Current Session Personalization

10
Local Index Personalization

• Retrieve the top-k documents from the history log
  that match the query
• Query expansion
• Augment query with
• All terms in the titles of the top-k documents
• The 20 terms with the highest BM25 score
• All terms from the summary of the top-k documents
• Re-rank results based on cosine similarity with
  the expanded query vector

11
Query travel to Indonesia
12
Query travel to Indonesia
13
Current Session Personalization

• Make use of current search context in order to
  determine intended meaning
• Associate with each query a user profile vector
• Update user profile with each result click
• Re-rank unseen results based on cosine similarity
  with the user profile vector

14
Query travel to Indonesia
15
Current Session Personalization

• For a new query
• Construct the user profile by augmenting the
  query vector with terms from similar browsed
  documents
• Re-rank results based on the cosine similarity
  with the user profile vector

16
Query java map
17
Further Considerations

• Enhance reranking procedure
• Make use of query chains
• Make use of click stream
• Run experimental tests to compare the new ranks
  to the original ranks