Learn from Web Search Logs to Organize Search Results

1
Learn from Web Search Logs to Organize Search
Results

• SIGIR2007
• Xuanhui Wang ChengXiang Zhai
• UIUC

2
Outline

• Motivation
• Problem
• Research Statement
• Related Work
• Solution
• Experiment
• Discussion
• Conclusion

3
Motivation

• Effective organization of search results is
  critical for improving the utility of any search
  engine.
• Allow a user to navigate into relevant documents
  quickly.

4
Problem

• Clustering search results
• Problem
• The clusters discovered do not correspond to the
  interesting aspects of a topic from the users
  perspectives.
• e.g., users are often interested in finding
  either phone codes or zip codes when entering
  the query area codes. The results are clustered
  into local codes and international codes.
• The cluster labels generated are not informative
  to allow a user to identify the right cluster.
• e.g., for the ambiguous query jaguar, a cluster
  may be labeled as panthera onca. Although this
  is an accurate label for a cluster with the
  animal sense of jaguar, if a user is not
  familiar with the phrase, the label would not be
  helpful.

5
Research Statement

• Purpose
• Learning interesting aspects of a topic from
  Web search logs and organizing search results
  accordingly.
• Car (car rental, car pricing)
• Generating more meaningful cluster labels using
  past query words entered by users.
• Jaguar (car, animal, software, and sports team)

6
Related Work

• Clustering search results
• Scatter/ Gather algorithm
• Grouper
• Suffix Tree Clustering (STC) algorithm
• Search logs
• Frequent Asked Questions
• Suggesting query substitutes
• Personalized search
• Learning retrieval ranking function

7
Search Engine Logs

• Sample

• All the titles, snippets, and URLs of the clicked
  Web pages of that query are used to represent the
  session.
• Aggregate all the sessions which contain exactly
  the same queries together.
• Pseudo-document

8
Solution

• Given a input query
• Get its related information from search engine
  logs. All the information forms a working set.
• Learn aspects from the information in the working
  set. Each aspect is labeled with a representative
  query.
• Categorize and organize the search results of the
  input query according to the aspects learned
  above.

9
Finding Related Past Queries

• History data set
• N pseudo-documents
• Use OKAPI to calculate the similarity between
  query q and Qi
• Given a query q, we use Hqd1,,dn to represent
  the top ranked pseudo-documents from the history
  collection H.

10
Learning Aspects by Clustering

• Star clustering algorithm
• Given Hq, it constructs a pair-wise similarity
  graph on this collection based on the vector
  space model.
• A similarity graph Gs can be constructed using a
  similarity threshold parameter s.
• The center provides a label for the cluster.

11
Categorizing Search Results

• A simple centroid-based method
• For each discovered aspect Ci, it builds a
  centroid prototype pi by
• For each search result sj, it computes the cosine
  similarity between sj and pi, and then it assign
  sj to the aspect with the highest score.

12
Data Collection

• MSN search log data set
• 8,144,000 queries, 3441,000 distinct queries, and
  4,649,000 distinct URLs in the raw data. (2/3,
  1/3)
• Only keep those frequent, well-formatted, English
  queries, it gets 169,057 unique queries.
• Obtain 172 and 177 test cases in the first and
  second test sets respectively.

13
Experiment

• Experiment Design
• Baseline default ranked list from a search
  engine
• Cluster-based simply clustering search results
• Log-based their approach
• Evaluations
• P_at_5
• MRR

14
Experimental Results
15
Detailed Analysis

• Diversity Analysis
• Diversity analysis
• Difficulty analysis

16
Parameter Setting
17
An Illustrative Example

• Query area codes (phone codes, zip codes)

18
Conclusion

• Use search engine logs to learn the interesting
  aspects, and then categorize the search results
  into the aspects learned.
• Their method can improve the ranking baseline,
  and generate more meaningful aspect labels.
• Future work
• Personal search log to improve the organization
  of search results for each individual user.