Topic-Sensitive PageRank

1
Topic-Sensitive PageRank

• Taher H. Haveliwala
• 2002

2
Abstract

• Targetimproving the ranking of search-query
  results
• Beforeusing the link structure of the Web, to
  capture the relative importance of Web pages,
  independent of any particular search query
• Nowa set of PageRank vectors, biased using a set
  of topics, to capture more accurately the notion
  of importance with respect to a particular topic

3
Abstract contribution

• more accurate rankings than generic PageRank
• Compute topic-sensitive PageRank scores for pages
  satisfying the query using the topic of the query
  keywords
• Considering searches done in context
• Compute the topic-sensitive PageRank scores using
  the topic of the context in which the query
  appeared

4
1. Introduction

• HITS 14
•  a link analysis algorithm
• Hubs
• Authorities
• Include content analyst 4
• Automatically compiling resource lists for
  general topics 8

5
1. Introduction - PageRank algorithm7,16

• rank vector - apriori importance -gt estimate
  pages on the Web
• Computed once
• Offline
• independent of the search query (con)
• importance scores are used in conjunction with
  query-specific IR scores to rank the query
  results

6
1. Introduction - Advantage of PageRank

• query-time cost of incorporating the precomputed
  PageRank importance score for a page is low
• PageRank is generated using the entire Web graph
  rather than a small subset

7
1. Introduction - Method in this paper

• allows the query to influence the link-based
  score(HITS)
• requires minimal query-time processing (PageRank)
• biased with a different topic

8
1. Introduction -making PageRank topic-sensitive

• avoid the problem of heavily linked pages getting
  highly ranked for queries(no particular
  authority)
• Hilltop 5 that is designed to improve results
  for popular queries
• Generates a query-specific authority score by
  detecting and indexing pages that appear to be
  good experts for certain keywords
• experts were not found will not be handled by the
  Hilltop algorithm.

9
1. Introduction -making PageRank topic-sensitive

• 17Propose using a set of Web Pages terms for
  influencing the computation.
• An approach for enhancing search rankings by
  generating a PageRank vector for each possible
  query term was recently proposed in 18 with
  favorable results
• requires considerable processing time and storage
• not easily extended to make use of user and query
  context

10
1. Introduction - two query scenarios

• Scenarios1assume a user with a specific
  information need issues a query
• Determine the topics most closely associated with
  the query, and use the appropriate
  topic-sensitive PageRank vectors for ranking the
  documents satisfying the query.

11
1. Introduction - two query scenarios

• Scenario2user is viewing a document (for
  instance, browsing the Web or reading email), and
  selects a term from the document for which he
  would like more information.

12
Summary of approach

• generate 16 topic-sensitive PageRank vectors
  using URLs from a top-level category from the
  Open Directory Project (ODP)
• At query time, calculate the similarity of the
  query to each topics
• take the linear combination of the
  topic-sensitive vectors, weighted using the
  similarities of the query to the topics
• link-based computations are performed offline,
  the query-time costs are not much

13
2. Review of PageRank

• page u link to page v
• Example
• Yahoo -gt important page(many pages point to it)
• pointed to from Yahoo! are probably important
• Nu -gt out degree of page u
• Rank(p) importance of page p
• link (u ,v) confers units of rank to v

14

• N is the number of pages , assign all pages the
  initial value 1/N
• Bv represent the set of pages pointing to v
• The final vector
• contains the PageRank vector over the Web
• computed only once after each crawl of the Web

15

• expressed as the following eigenvector
  calculation
• M -gt square stochastic matrix corresponding to
  the directed graph G of the Web
• Page j to Page I , mij 1/Nj
• Repeatedly multiplying Rank by M yields the
  dominant eigenvector Rank of the matrix M

16
An example
v2
v1
v3
v5
v4
17

• PageRank can be viewed as the stationary
  probability distribution over pages induced by a
  random walk on the Web
• To convergence - M is irreducible and aperiodic
• Dumping factor1 a to restrict rank sink
• add transition edges of probability a /N between
  every pair of nodes in G

18
Rank Sink

• Page A points to Page B and Page B points to Page
  A, and the PageRank value for these pages
  increases

RefAnalysis of Rank Sink Problem in PageRank
Algorithm
19
The key to creating topic-sensitive PageRank is
to bias the computation to increase the effect of
certain categories of pages by using a nonuniform
Nx1 personalization vector for
20
3. Topic-Sensitive PageRank - Method in this
article

• precompute the multiple importance scores for
  each page
• a set of scores of the importance of a page with
  respect to various topics
• combined to form a composite PageRank score
• to produce the final rank of the query

21
3. Topic-Sensitive Pagerank - 3.2 ODP biasing

• To generate a set of biased PageRank vectors
  using a set of a basis topics.
• Performed once
• Offline
• Personalization vector
• 16 different biased PageRank vectors (using 16
  top-level of ODP)
• Tj set of URLs in the ODP category cj

DMOZ Open Directory Project 16 top-level topics
22
3.2 Personalization vector
23
3.3 Query Time Importance Score

• The second is performed at query time
• Given a query q, let q be the context of q.
• Let qi be the ith term in the query context
• Then given the query q, compute for each cj the
  following
• reflects the interests of user k

24

• the query-sensitive importance score of each of
  these retrieved URLs
• rankjd is the rank of document d given by the
  rank vector
• The results are ranked according to this
  composite score sqd

25

• random surfer modelvisits a web page with a
  certain probability which derives from the page's
  PageRank
• wj is the coefficient used to weight the jth rank
  vector
• With probability 1- a a random surfer on page u
  follows an outlink of u

26
4. Experimental Results

• A series of experiments
• 4-1 describe the similarity measure use to
  compare two rankings
• 4-2 investigate how the induced rankings vary,
  based on both the topic used to bias the rank
  vectors
• 4-3 the retrieval performance of ordinary
  PageRank versus topic-sensitive PageRank.
• 4-4 how the use of query context can be used in
  conjunction with topic-sensitive PageRank

27
4. Experiment Data

1. crawl contained roughly 280,000 of the 3 million
   URLs in the ODP.
2. 35 queries in paper 9 show at Table1

28
4-1 Similarity Measure - First measure

• First measure
• degree of overlap
  between the top n URLs of two rankings
• n 20, use to compare
• it does not indicate the degree to which the
  relative orderings of the top n URLs of two
  rankings

29
4-1 Similarity Measure - second measure
Kendalls distance measure9

• KSim(T1,T2) is the probability that and
  agree on the relative ordering of a randomly
  selected pair of distinct nodes
• U union of the URLs in and

30

• Ref https//en.wikipedia.org/wiki/Kendall_tau_dis
  tance

31
4.2 Effect of ODP Biasing

• bias factor a
• affects the degree to which the resultant vector
  is biased towards the topic vector used for
• For a 1, the URLs in the bias set Tj will be
  assigned the score
• as a 0, the content of Tj becomes irrelevant to
  the final score

32

• Use a 0.25(heuristically)
• the induced rankings of query results are not
  very sensitive to the choice of a
• The average overlap
• between the top 20
• results for the two
• values of is very high

33

• differences across different topically-biased
  PageRank vectors is much higher

34

• investigate which of these rankings is best for
  specific queries
• Table 5 shows the top 5 ranked URLs

35
4.3 query-sensitive scoring

• how effectively we can utilize the ranking
  precision
• intuitively the most relevant
  categories for the query
• Use only the top three highest values categories
  to compute sqd score

36
4.3 query-sensitive scoring
37
4.3 query-sensitive scoring -experiment

• To compare the query-sensitive approach to
  PageRank
• 10 queries
• 5 volunteers
• Each query, the volunteer
• was shown 2 result rankings
• Top 10 results with the unbiased PageRank vector
• Top 10 results with the composite sqd score
• Select all URLs relevant to the query
• Choose the better ranking results

38
4.3 query-sensitive scoring -result
39
4.4 context-sensitive scoring

• Using the context can help disambiguate the query
  term and yield results that more closely reflect

40
(No Transcript)
41
5.Sources of search context

1. the history of queries issued leading up to the
   current query is another form of query context
2. Jordan and basketball
3. sort of hierarchical directory
4. User context
5. Browsings patterns
6. Bookmarks
7. Email archive

42
6. Ongoing Work

• discovering sources of search context
• development of the best set of the basis
  topics(second of third level of Open Directory
  hierarchy) -gt efficiency problem
• Creating the dumping vector to create the topic
  sensitive rank vectors -gt being more resistant