Combating Web Spam with TrustRank

1
Combating Web Spam with TrustRank

• Presented by
• Nitin Mittal

2
About the Paper

• Name
• Combating Web Spam with TrustRank
• Authors
• Zoltan Gyongyi
• Hector Garcia-Molina
• Jan Pedersen
• Year
• 2004
• Topic
• This paper presented the first attempt to
  formalizing the problem of combating Web Spam
  and introduced a comprehensive solution to
  assist in the detection of Web Spam.

3
Presentation Outline

• Introduction
• Preliminaries
• Trust Rank Algorithm
• Selecting Seeds set
• Experiment Result
• Conclusions
• Further improvements

4
Introduction

• Web Spam
• The term web Spam refers to hyperlinked pages
  on the World Wide Web that are created with the
  intention of misleading search engines.
• Techniques of Web Spam
• By adding the thousands of keywords to home page,
  so that a search engine will index the key words
  and return the bogus site as an answer to queries
  on that keywords.
• Creating a large number of bogus web pages, all
  pointing to a single target page. Since many
  search engines ranks the pages with the
  consideration on incoming links.

5
Introduction(Cont.)

• Issues with Web Spam
• Its hard to check if pages has the contents
  related to the keywords listed at its main page
  or they are just inserted for Web Spam.
• Inter linking websites may represent useful
  relations between the sites, or they may have
  been created to boost the rank of each others
  pages.
• Not an easy task for computer to categories web
  pages due to the large number of web pages.

6
Introduction(Cont.)

• Contribution of this paper are
• Formalize the problem of Web Spam and Spam
  detection algorithms.
• Defined metrics for assessing the efficacy of
  detection algorithms.
• Presented schemes for selecting seed sets of
  pages to be manually evaluate.
• Introduced the TrustRank algorithms for
  determining the likelihood that pages are
  reputable(not Spam).
• Discuss results of an extensive evaluation based
  on 31 million sites crawled by the Alta Vista
  search engine.

7
Preliminaries

• Web Model
• Page Rank

8
Preliminaries (Cont)

• 1. Web Model
• Web is modeled as a graph G(V,E), Consisting
  of a set V of N pages (vertices) and a set E of
  directed links (Edges) that connect pages.
• Multiple hyperlink between two pages p and q, is
  collapsed into a single link(p,q) ? E.
• In-degree, number of inlinks of a page p, is
  l(p)
• Out-degree, number of outlinks of a page p, is
  ?(p)
• Pages without outlinks are referred to as
  non-referencing pages.
• Pages without inlinks are referred to as
  unreferenced pages.
• Pages that are both unreferenced and
  non-referencing at the same time are referred as
  isolated pages.

9
Preliminaries (Cont)

• Page 1 is unrefernced page.
• Page 4 is non-referncing page.

10
Preliminaries (Cont)

• Transposition matrix, T
• Inverse Transposition matrix, U

11
Preliminaries (Cont)

• 2. PageRank
• PageRank is a well known algorithm that uses
  link information to assign global importance
  scores to all pages on the web.
• The intuition behind PageRank is that a web page
  is important, if several other important web
  pages point to it.
• The PageRank score r(p) of a page p is defined
  as
• Where ? is a decay factor.

12
TrustRank Algorithm

• 1. Assessing Trust
• 2. Computing Trust
• 3. The TrustRank Algorithm

13
TrustRank Algorithm(Cont)

• 1.Assessing Trust
• Determination of initial set of pages as if it
  is a Web Spam or not, requires human evaluation.
  A notion is introduced as human checking a page
  for spam by a binary oracle function O over all
  pages p?V
• Oracle invocations are expensive and time
  consuming. Thus, our objective is to call
  function O on selective pages.

14
TrustRank Algorithm(Cont)

• To discover good pages without invoking the
  oracle function on the entire web, paper made an
  important empirical observation that Good pages
  seldom points to bad ones.
• Spam pages can, and in fact often do, link to
  good pages.

15
TrustRank Algorithm(Cont)

• Threshold Trust Property
• If a page p receives a score above ? then we
  know that it is good. Otherwise, we cannot tell
  anything about p.
• Where, T(p) is a Ideal Trust Property.

16
TrustRank Algorithm(Cont)

• 2. Computing Trust
• There is a limited budget L, of function O
  invocation. The subsets of good and bad seed
  pages by ? and ?-, respectively. The remaining
  pages are not checked by the human expert, we
  assign them a trust score of ½ to signal out lack
  of information.

17
TrustRank Algorithm(Cont)

• Trust Propagation
• As we are not sure that pages reachable from
  good seeds are indeed good.
• The further away we are from good seed pages,
  the certain we are that a page is good.
• Trust Attenuation
• Trust Dampening

18
TrustRank Algorithm(Cont)
Trust Splitting
19
TrustRank Algorithm(Cont)
20
TrustRank Algorithm(Cont)

• Step 1.
• S0.08,0.13,0.08,0.10,0.09,0.06,0.02
• Step 2.
• ?2,4,5,1,3,6,7
• Step 3.
• Invoke Oracle function on set s2,4,5.
• Step 4.
• d0,1/2,0,1/2,0,0,0
• Step 5.
• Evaluate the TrustRank score on whole set.

21
Selecting Seed Set

• 1. Inverse PageRank
• We could select select seed pages based on the
  number of outlinks.
• Inverse PageRank is a heuristic, that is works
  well in practice.
• 2. High PageRank
• As high PageRank pages are likely to point to
  other high-PageRank pages, then good trust scores
  will also be propagated to pages that are likely
  to be at the top of result sets.

22
Experiment Result

• In August 2003, using the complete set of pages
  crawled and indexed by the Alta Vista search
  engine, they group several billion pages into
  31,003,946 sites.
• The author of this paper played the role of the
  oracle.
• After conducting experiments to compare the
  inverse PageRank and the high PageRank seed
  selection schemes,Inverse PageRank was selected.
• With the help of major web directories, out of
  top 25,000 websites, they selected 7,900.
• Oracle is called on top 1250, and selected 178
  sites as good seeds.

23
Experiment Result (Cont.)

PageRank Versus TrustRank
24
Experiment Result (Cont.)

• PageRank Versus TrustRank
• PageRank algorithm does not incorporate any
  knowledge about the quality of a site, nor does
  it explicitly penalize badness. Where as,
  TrustRank is meant to differentiate good and bad
  sites.
• Almost no spam in top 5 TrustRank buckets, while
  it is surprising that almost 20 of the second
  PageRank bucket is bad.
• Spam is highest in PageRank buckets 9 and 10,
  while corresponding TrustRank buckets 9 has
  different values.

25
Experiment Result (Cont.)

• Other Strategies to evaluate results
• 1.Pairwise Orderedness
• 2.Precision and Recall

26
Experiment Result (Cont.)

• 1. Pairwise Orderedness
• Pairwise Orderedness is related to ordered
  trust property, which tell us that what fraction
  of the pairs of website (p,q), for which
  T(Threshold Trust Property) does not make
  mistake.
• TrustRank constantly outperforms both the
  ignorant function and PageRank.

27
Experiment Result (Cont.)

• 2. Precision and Recall
• Precision-
• It is the fraction of good among all pages in
  the set X that have a trust score above the
  average trust score.
• Recall-
• It is the ratio between the number of good pages
  with a trust score above average trust score and
  the total number of good pages in the set X.

28
Conclusions

• Search engines combats Web Spam with a variety of
  ad hoc, often proprietary techniques.
• Paper presented a solution to assist in the
  detection of web Spam with the help of Trust
  Rank.
• Results shows that presented solution can
  effectively identify a significant number of
  strongly reputable (non-Spam) pages.
• Trust Rank can be used either separately to
  filter the index, or in combination with PageRank
  and other metrics to rank search results.

29
Limitation

• Although, TrustRank guarantees that top-scored
  sites are good one, and has better result then
  PageRank. But, TrustRank is unable to effectively
  separate low-scored good sites from bad one, due
  to the lack of distinguishing features (inlinks)
  of the sites.

30
Further improvements

• Explore the interplay between dampening and
  splitting for trust propagation.
• Instead of selecting the entire seed set at once,
  an iterative process could be implemented to
  reconsider the pages that oracle should evaluate
  next, based on the previous outcome of oracle
  results, after oracle has evaluated some pages.

31

• Thank You