Personalizing Search via Automated Analysis of Interests and Activities

1
Personalizing Search via Automated Analysis of
Interests and Activities

• Jaime Teevan, MIT, CSAIL
• Susan, MS Research
• Eric, MS Research

2
Outline

• Introduction
• Pursuit of Personalization
• Corpus, User, Doc Query Representation
• Evaluation Framework
• Results
• Conclusion Future Work

3
Introduction

• Web search for IR returns a wide range of
  results etc.
• Beyond the commons Investigating the value of
  personalizing Web search. (workshop on new
  technologies for PIA 2005)
• They found that there is an opportunity to
  achieve significant improvement by
  custom-tailoring search results to individuals,
  and were thus motivated to pursue search
  algorithms that return personalized results
  instead of treating all users the same.

4
Pursuit of Personalization

• Focused on re-ranking the top research results
  locally.
• They explored Web search personalization by
  modifying BM25 as their Probabilistic weighting
  scheme.
• Corpus representation N, ni
• User representation R, ri
• Doc Query representation terms

5
Corpus, User, Doc Query Representation

• Corpus Representation
• Full-text VS Title Snippet
• User Representation
• A rich index of personal content that captured a
  users interests and computational activities
• Within last month VS Full index of documents
• Doc Query Representation
• Full-text VS Title Snippet

6
Evaluation Framework

• 15 participants, top 50 search results.
• For 10 self-selected Queries
• 2 way to evaluate queries
• Collected a total of 131 queries
• DCG (Discounted Cumulative Gain)
• Rank(D1,..Dn), Relevance(3,2,3,0,0,1,2,)

7
Results

• One-way effects in which we hold all but one
  variable constant.
• Baseline Comparisons
• Combining Rankings

8
One-way effects

• F2

9
Best combination of parameters for personalized
search

• Corpus Representation Approximated by the
  result set title and snippets.
• User Representation Built from the users
  entire personal index.
• Document Query Representation Documents
  represented by the title and snippet returned by
  the search engine, with query expansion based on
  words that occur near the query term.

10
Baseline Comparisons

• F3

11
Combining Rankings
12
Conclusion

• They have investigated the feasibility of
  personalizing Web search by using an
  automatically constructed user profile as
  relevance feedback in their ranking algorithm.
• Better than explicit relevance feedback.
• Its possible to approximate the corpus, making
  efficient client-side computation feasible.

13
Future Work