Forschungszentrum L3S

1
Site Level Noise Removal for Search
Engines Written By Andre Carvalho, Paul
Alexandru Chirita, Edleno de Moura Pavel Calado,
Wolfgang Nejdl
2
Outline

• Motivation
• Noisy Web Structures at the Site Level
• Mutual Site Reinforcement Relations
• Abnormal Site Support
• Site Level Link Alliances
• Experiments
• Conclusions

3
Outline

• Motivation
• Noisy Web Structures at the Site Level
• Mutual Site Reinforcement Relations
• Abnormal Site Support
• Site Level Link Alliances
• Experiments
• Conclusions

4
Motivation

• Web Spam has become an industry
• All previous approaches to link spam detection
  use various page level algorithms
• Spammers are now creating more complex PageRank
  boosting graphs, across several sites
• There is more than just Spam
• Noisy links are hyperlinks created for any
  purpose, but that of expressing a vote for the
  target page
• Each branch of a company has its own site, these
  sites being interconnected by a navigational
  structure
• Replicated sites pointing towards the newer
  replica

5
Outline

• Motivation
• Noisy Web Structures at the Site Level
• Mutual Site Reinforcement Relations
• Abnormal Site Support
• Site Level Link Alliances
• Experiments
• Conclusions

6
Mutual Site Reinforcement

• Characterized by many link exchanges between two
  or more sites (see figure)
• Several counting approaches are possible
• Only the bidirectional link exchanges (3)
• Link density, i.e., all links between sites, as
  within a unidirectional counting (9)
• Minimum no. of inter-site links S ? S and S ? S
  (4, work in progress)

7
Mutual Site Reinforcement and Ranking

• Detection algorithm
• For each page p from S, take all q from S other
  than S, out-neighbors of p
• If p is also out-neighbor of q, then increase
  BMSR (bidirectional link)
• Increase UMSR always (unidirectional link)
• Inclusion in PageRank is then straightforward
• If the amount of link exchanges of either type
  exceeds a give threshold, then remove all links
  between S and S
• Another solution would be to downgrade these
  links instead (not tested)
• Thresholds
• 250 for UMSR (tried quite several values between
  10 and 300)
• 2 for BMSR (tried up to 4) shows that incorrect
  votes are usually pair-wise

8
Outline

• Motivation
• Noisy Web Structures at the Site Level
• Mutual Site Reinforcement Relations
• Abnormal Site Support
• Site Level Link Alliances
• Experiments
• Conclusions

9
Abnormal Site Support

• Based on the idea that
• For a site S, there should not be another site
  S, whose number of links towards S is above a
  certain percentage of the total number of links S
  receives overall
• Best threshold was 2 (tried value ranging from
  0.5 up to 20)
• Algorithm
• For all pairs of sites (S, S), calculate
• Support Links (S ? S) / In-links (S)
• If Support gt e then remove all links S ? S

10
Outline

• Motivation
• Noisy Web Structures at the Site Level
• Mutual Site Reinforcement Relations
• Abnormal Site Support
• Site Level Link Alliances
• Experiments
• Conclusions

11
Site Level Link Alliances

• Most complex and costly structures
• If a page p has in-links from pages
• i1, i2, . . ., in, and
• these latter pages are highly connected,
• Then they are suspect of being part of a
  structure which could deceive popularity
  ranking algorithms
• It is a Site Level approach only in the sense
  that the in-linking pages usually belong to
  different sites
• Solution
• Each page is downgraded according to its
  susceptivity of being part of such a malicious
  structure

12
Site Level Link Alliances and Susceptivity

• First step involves computing the Maliciousness
  Susceptivity of each page p
• Let Tot be the number of out-links of all pages q
  linking to p
• and TotIn be the number of out-links of all
  pages q linking to p, such that they point to
  some other page linking to p
• Susceptivity is then the ratio between TotIn and
  Tot
• For p this is 22 / 26

13
Site Level Link Alliances and Ranking

• Simply removing all links to pages with a high
  susceptivity would be too harsh on them (they
  would be practically classified as spam)
• Use a downgrading method instead

14
Outline

• Motivation
• Noisy Web Structures at the Site Level
• Mutual Site Reinforcement Relations
• Abnormal Site Support
• Site Level Link Alliances
• Experiments
• Conclusions

15
Experimental Setup Dataset

• TodoBR Search Engine http//www.todobr.com.br/
• About 12 M pages, query log with about 11 M
  entries
• Extracted queries from the log, as follows
• Bookmark queries, seeking for a specific page
  (100)
• Topic queries, seeking for information on a given
  topic (60)
• Each set of queries contained two equal subsets
• Popular queries, i.e., most frequent Show how
  many of the top sites have been downgraded by our
  algorithms
• Randomly selected queries Show the impact on
  pages with a low PageRank

16
Experimental Setup Evaluation

• 14 undergraduate and graduate CS students
• Bookmark queries evaluated using MRR
• and MEANPOS, i.e., the average position of the
  first relevant answer
• (more sensible to low-rank relevant results,
  e.g., places 15 and 40)
• Topic queries used P_at_5, P_at_10, and MAP also,
  results were categorized into non-relevant (0),
  relevant (1), and highly relevant (2)
• Web sites simply defined as the host name part of
  the URL
• Fine-tuned the thresholds using the MRR
  experiments

17
Overall Results, Bookmark Queries
18
Overall Results, Topic Queries
19
Practical Issues

• Amount of links removed (where applicable)
• SLLA SLAbS could thus be preferred to increase
  computation speed, as BMSR does not remove too
  many links
• Let V the no. of nodes, E the no. of edges, M the
  average out-degree, and P the average in-degree.
  Then, the complexities are
• O(E) for UMSR and O(VMlog(M)) for BMSR
• O(VP) for SLAbS
• O(VM2log(M)) for SLLA
• All algorithms are trivially parallelizable

20
Outline

• Motivation
• Noisy Web Structures at the Site Level
• Mutual Site Reinforcement Relations
• Abnormal Site Support
• Site Level Link Alliances
• Experiments
• Conclusions

21
Conclusions and Further Work

• First approach to detect link noise at the site
  level
• Three algorithms
• Mutual Site Reinforcement Relations
• Site Level Abnormal Support
• Site Link Alliances
• Improvement of
• 26.98 in MRR for popular bookmark queries,
• 20.92 for randomly selected bookmark queries,
• and up to 59.16 in Mean Average Precision for
  topic queries
• Identified up to 16.7 of the links from our
  collection as noisy
• To do
• Try more complex site definitions, as well as new
  algorithms
• Automatic parameter tuning (hard)

22
Thank You!