Survey on web spam detection

15 Pages·2012·0.38 MB·English

Survey on Web Spam Detection: Principles and Algorithms
Nikita Spirin Jiawei Han
DepartmentofComputerScience DepartmentofComputerScience
UniversityofIllinoisatUrbana-Champaign UniversityofIllinoisatUrbana-Champaign
Urbana,IL61801,USA Urbana,IL61801,USA
[email protected] [email protected]
ABSTRACT as means of malware and adult content dissemination and
fishing attacks. For instance, [39] ranked 100 million pages
Searchenginesbecameadefactoplacetostartinformation
usingPageRankalgorithm[86]andfoundthat11outoftop
acquisition on the Web. Though due to web spam phe-
20 results were pornographic websites, that achieved high
nomenon, search results are not always as good as desired.
ranking due to content and web link manipulation. Last, it
Moreover, spam evolves that makes the problem of provid-
forcesasearchenginecompanytowasteasignificantamount
ing high quality search even more challenging. Over the
of computational and storage resources. In 2005 the total
lastdecaderesearchonadversarialinformationretrievalhas
worldwide financial losses caused by spam were estimated
gainedalotofinterestbothfromacademiaandindustry. In
at $50 billion [63], in 2009 the same value was estimated
this paper we present a systematic review of web spam de-
already at $130 billion [64]. Among new challenges which
tection techniques with the focus on algorithms and under-
are emerging constantly one can highlight a rapid growth
lying principles. We categorize all existing algorithms into
of the Web and its heterogeneity, simplification of content
three categories based on the type of information they use:
creationtools(e.g.,freewebwikis,bloggingplatforms,etc.)
content-based methods, link-based methods, and methods
anddecreaseinwebsitemaintenancecost(e.g.,domainreg-
basedonnon-traditionaldatasuchasuserbehaviour,clicks,
istration, hosting, etc.), evolution of spam itself and hence
HTTP sessions. In turn, we perform a subcategorization of
appearance of new web spam strains that cannot be cap-
link-basedcategoryintofivegroupsbasedonideasandprin-
tured by previously successful methods.
ciples used: labels propagation, link pruning and reweight-
Web spam phenomenon mainly takes place due to the fol-
ing, labels refinement, graph regularization, and feature-
lowing fact. The fraction of web page referrals that come
based. Wealsodefinetheconceptofwebspamnumerically
from search engines is significant and, moreover, users tend
and provide a brief survey on various spam forms. Finally,
to examine only top ranked results. Thus, [98] showed that
we summarize the observations and underlying principles
for85%ofthequeriesonlythefirstresultpageisrequested
applied for web spam detection.
and onlythefirst threeto fivelinksareclicked [65]. There-
fore, inclusion in the first SERP1 has a clear economic in-
Keywords
centiveduetoanincreaseinwebsitetraffic. Toachievethis
goal website owners attempt to manipulate search engine
web spam detection, content spam, link spam, cloaking,
rankings. This manipulation can take various forms such
collusion, link farm, pagerank, random walk, classification,
as the addition of a surrogate content on a page, excessive
clustering,websearch,userbehaviour,graphregularization,
and undeserved link creation, cloaking, click fraud, and tag
labels propagation
spam. We define these concepts in Section 2, following the
work [54; 82; 13; 57]. Generally speaking, web spam man-
1. INTRODUCTION
ifests itself as a web content generated deliberately for the
Spampervadesanyinformationsystem,beite-mailorweb, purpose of triggering unjustifiably favourable relevance or
social, blog or reviews platform. The concept of web spam importance of some web page or pages [54]. It is worth
or spamdexing was first introduced in 1996 [31] and soon mentioning that the necessity of dealing with the malicious
was recognized as one of the key challenges for search en- contentinacorpusisakeydistinctivefeatureofadversarial
gine industry [57]. Recently [50; 97], all major search en- informationretrieval[41]incomparisonwiththetraditional
gine companies have identified adversarial information re- information retrieval, where algorithms operate on a clean
trieval [41] as a top priority because of multiple negative benchmark data set or in an intranet of a corporation.
effectscausedbyspamandappearanceofnewchallengesin Accordingtovariousstudies[85;25;12]theamountofweb
this area of research. First, spam deteriorates the quality spam varies from 6 to 22 percent, which demonstrates the
ofsearchresultsanddepriveslegitimatewebsitesofrevenue scope of the problem and suggests that solutions that re-
that they might earn in the absence of spam. Second, it quire manual intervention will not scale. Specifically, [7]
weakenstrustofauserinasearchengineproviderwhichis shows that 6% of English language web pages were classi-
especially tangible issue due to zero cost of switching from fied as spam, [25] reports 22% of spam on a host level and
one search provider to another. Third, spam websites serve [12] estimates it as 16.5%. Another group of researchers
study not only the cumulative amount of spam but its dis-
1Search Engine Result Page.
SIGKDD Explorations Volume13,Issue 2 Page50
tributionamongcountriesandtopleveldomains[85]. They • Body Spamming. In this case the body of a page is
report 13.8% of spam in the English speaking internet, 9% modified. This is the most common form of content
inJapanese,22%inGerman,and25%inFrench. Theyalso spam because it is cheap and simultaneously allows
showthat70%ofpagesinthe*.bizdomainand20%inthe applying various strategies. For instance, if a spam-
*.com domain are spam. merwantstoachieveahighrankingofapageforonly
This survey has two goals: first, it aims to draw a clear limitedpredefinedsetofqueries,theycanusetherep-
roadmap of algorithms, principles and ideas used for web etition strategy by overstuffing body of a page with
spamdetection;second,itaimstobuildawarenessandstim- termsthatappearinthesetofqueries(thereevenwas
ulatefurtherresearchintheareaofadversarialinformation aspammingcompetitiontorankhighestforthequery
retrieval. To the best of our knowledge there is no com- “nigritude ultramarine” [38]). On the other hand, if
prehensive but concise web spam mining survey with the the goal is to cover as many queries as possible, the
focus on algorithms yet. This work complements existing strategycouldbetousealotofrandomwordsatonce.
surveys [54; 13; 82; 57] and a book [23] on the topic of web To hide part of a content used to boost ranking, site
spam. [54] presents a web spam taxonomy and provides owners make it coloured in the same way as a back-
abroadcoverageofvariouswebspamformsanddefinitions. ground so that only machines can recognize it.
In [13] issues specific to web archives are considered. [82;
• Meta-Tags Spamming. Because meta-tags play a spe-
57]enumeratespamformsanddiscusschallengescausedby
cificroleinadocumentdescription,searchenginesan-
spam phenomenon for web search engines. [23] is a recent
alyze them carefully. Hence, the placement of spam
book, which provides the broadest coverage of web spam
content in this field might be very prospective from
detection research available so far. We think that paral-
spammer point of view. Because of the heavy spam-
lel research on email spam fighting [19; 95; 113] and spam
ming, nowadays search engines give a low priority to
on social websites [58] might also be relevant. [33] presents
this field or even ignore it completely.
a general framework for adversarial classification and ap-
proaches the problem from game-theoretical perspective.
• Anchor Text Spamming. The usefulness of anchor
The paper is organized as follows. In Section 2 we pro- text3forwebrankingwasfirstintroducedin1994[80],
vide a brief overview of web spam forms following [54; 82;
andinstantlyspammersaddedthecorrespondingstrat-
13; 57]. Then we turn to content-based mining methods
egytotheir“arsenal”. Spammerscreatelinkswiththe
in Section 3.1. In Section 3.2 we provide a careful cover-
desired anchor text (often unrelated to a linking page
ageoflink-basedspamdetectionalgorithms. InSection3.3
content) in order to get the “right” terms for a target
we consider approaches to fight against spam using click-
page.
through and user behaviour data, and by performing real-
time HTTP sessions analysis. Finally, we summarize key • URL Spamming. Some search engines also consider a
principles underlying web spam mining algorithms in Sec- tokenized URL of a page as a zone. And hence spam-
tion 4 and conclude in Section 5. merscreateaURLforapagefromwordswhichshould
be mentioned in a targeted set of queries. For exam-
ple,ifaspammerwantstoberankedhighforthequery
2. WEBSPAMTAXONOMY
“cheap acer laptops”, they can create a URL
“cheap-laptops.com/cheap-laptops/acer-laptops.html”4.
2.1 ContentSpam
Thecontentspamisprobablythefirstandmostwidespread With the advent of link-based ranking algorithms [86; 68]
form of web spam. It is so widespread because of the fact content spam phenomenon was partially overcome. How-
that search engines use information retrieval models based ever,spamisconstantlyevolvingandsoonafterwardsspam-
onapagecontenttorankwebpages,suchasavectorspace mersstartedconstructinglinkfarms[53;3],groupsofhighly
model [96], BM25 [94], or statistical language models [114]. interconnected spam pages, with the aim to boost ranking
Thus,spammersanalyzetheweaknessesofthesemodelsand of one or a small number of target pages in a farm. This
exploit them. For instance, if we consider TFIDF scoring form of spam is referred to as link spam.
(cid:2)
TFIDF(q,p)= TF(t)·IDF(t), (1) 2.2 LinkSpam
t∈q∧t∈p There are two major categories of link spam: outgoing link
spam and incoming link spam.
where q is a query, p is a page, and t is a term, then spam-
mers can try to boost TF of terms appearing on a page2. 2.2.1 Outgoinglinkspam
There are multiple facets based on which we can categorize
This is the easiest and cheapest method of link spam be-
content spam.
cause, first, a spammer have a direct access to his pages
Taking a document structure into account there are 5 sub-
and therefore can add any items to them, and second, they
types of content spamming. caneasilycopytheentirewebcataloguesuchasDMOZ5 or
Yahoo! Directory6 and therefore quickly create a large set
• Title Spamming. Due to high importance of the title
field for information retrieval [79] spammers have a 3a visible caption for a hyperlink
clear incentive to overstuff it so as to achieve higher 4itisworthnotingthattheexamplealsodemonstratesidea
overall ranking. of keyword repetition, because the word “laptops” appears
three times in a tokenized URL representation
2we assume that IDF of a term cannot be manipulated by 5dmoz.org
a spammer due to an enormous size of the Web 6http://dir.yahoo.com/
SIGKDD Explorations Volume13,Issue 2 Page51
of authoritative links, accumulating relevance score. Out- page refresh time to zero and initialize a refresh URL
going link spamming targets mostly HITS (Section 3.2.1) attributewithaURLofatargetpage. Moresophisti-
algorithm [68] with the aim to get high hub score. cated approach is to use page level scripts that aren’t
usuallyexecutedbycrawlersandhencemoreeffective
2.2.2 Incominglinkspam from spammers point of view.
InthiscasespammerstrytoraiseaPageRank(Section3.2.1)
• Accessible Pages. Thesearepageswhichspammers
score of a page (often referred to as a target page) or sim-
can modify but don’t own. For instance, it can be
plyboostanumberofincominglinks. Onecanidentifythe
Wikipedia pages, blog with public comments, a pub-
following strategies depending on an access to pages.
lic discussion group, or even an open user-maintained
• OwnPages. Inthiscaseaspammerhasadirectcon- web directory. Spammers exploit the opportunity to
be able to slightly modify external pages by creating
trol over all the pages and can be very flexible in his
strategies. Hecancreatehisownlinkfarm7 andcare- linkstotheirownpages. Itisworthnotingthatthese
strategies are usually combined. Thus, while spam-
fully tune its topology to guarantee the desired prop-
ming comments, adversaries can apply both link and
erties and optimality. One common link farm has a
anchor text spamming techniques.
topologydepictedonafig.1andisnamedasahoney-
pot farm. Inthiscaseaspammercreatesapagewhich
2.3 CloakingandRedirection
looks absolutely innocent and may be even authorita-
tive (though it is much more expensive), but links to Cloakingisthewaytoprovidedifferentversionsofapageto
thespammer’stargetpages. Inthiscaseanorganically crawlers and users based on information contained in a re-
aggregated PageRank (authority) score is propagated quest. Ifusedwithgoodmotivation,itcanevenhelpsearch
further to target pages and allows them to be ranked engine companies because in this case they don’t need to
higher. Moreaggressiveformofahoneypotschemais parse a page in order to separate the core content from a
hijacking, when spammers first hack a reputable web- noisy one (advertisements, navigational elements, rich GUI
siteandthenuseitasapartoftheirlinkfarm. Thus, elements). However, if exploited by spammers, cloaking
in 2006 a website for prospective CS students was hi- takes an abusive form. In this case adversary site owners
jackedandspammedwithlinkofpornographicnature. servedifferentcopiesofapagetoacrawlerandauserwith
the goal to deceive the former [28; 108; 110; 75]. For exam-
ple,asurrogatepagecanbeservedtothecrawlertomanipu-
lateranking,whileusersareservedwithauser-orientedver-
sionofapage. Todistinguishusersfromcrawlersspammers
analyze a user-agent field of HTTP request and keep track
of IP addresses used by search engine crawlers. The other
strategyistoredirectuserstomaliciouspagesbyexecuting
JavaScript activated by page onLoad() event or timer. It
isworthmentioningthatJavaScriptredirectionspamisthe
most widespread and difficult to detect by crawlers, since
Figure 1: Scheme of a typical link farm.
mostly crawlers are script-agnostic [29].
Spammerscanalsocolludebyparticipatinginlinkex- 2.4 ClickSpam
changeschemesinordertoachievehigherscale,higher
Since search engines use click stream data as an implicit
in-link counts, or other goals. Motivations of spam-
feedback to tune ranking functions, spammers are eager to
merstocolludearecarefullyanalyzedin[53]. Optimal
generate fraudulent clicks with the intention to bias those
properties of link farms are analyzed in [53; 3; 117].
functionstowardstheirwebsites. Toachievethisgoalspam-
Toreducetimespentonalinkfarmpromotionspam- mers submit queries to a search engine and then click on
mersarealsoeagertobuyexpiredandabandoneddo- linkspointingtotheirtargetpages[92;37]. Tohideanoma-
mainnames. Theyareguidedbytheprinciplethatdue lous behaviour they deploy click scripts on multiple ma-
to the non-instant update of an index and recrawling, chinesoreveninlargebotnets[34;88]. Theotherincentive
searchenginesbelievethatadomainisstillunderthe ofspammerstogeneratefraudulentclickscomesfromonline
control of a good website owner and therefore spam- advertising [60]. In this case, in reverse, spammers click on
mers can benefit with “resources” and reputation left adsofcompetitorsinordertodecreasetheirbudgets,make
by the previous website for some time. them zero, and place the ads on the same spot.
Wealsoconsiderredirectionasaninstanttypeofhon-
eypot scheme8. Here the spamming scheme works as 3. ALGORITHMS
follows. First, a honeypot page achieves high ranking
All the algorithms proposed to date can be roughly cate-
inaSERPbyboostingtechniques. Butwhenthepage
gorized into three groups. The first one consists of tech-
isrequestedbyauser,theydon’tactuallyseeit,they
niques which analyze content features, such as word counts
getredirectedtoatargetpage. Therearevariousways
or language models [44; 85; 100; 81; 101; 89; 40], and con-
toachieveredirection. Theeasiestapproachistoseta
tent duplication [42; 43; 103]. Another group of algorithms
7now link farm maintenance is in dozens of times cheaper utilizeslink-basedinformationsuchasneighbourgraphcon-
than before and the price is constantly decreasing nectivity[25;47;49;48;119;118;116],performslink-based
8some researchers consider it as a separate form of spam- trust and distrust propagation [86; 55; 71; 99; 112; 12; 52;
ming technique 26; 66; 11; 5; 8], link pruning [16; 84; 73; 93; 74; 35; 109;
SIGKDD Explorations Volume13,Issue 2 Page52
32; 111; 115; 6], graph-based label smoothing [121; 1; 30], only cheap-to-compute content features. They also showed
andstudystatisticalanomalies[4;7;38;9]. Finally,thelast thatcomputationallydemandingandglobalfeatures,forin-
groupincludesalgorithmsthatexploitclickstreamdata[92; stance PageRank, yield only negligible additional increase
37; 60] and user behaviour data [77; 78], query popularity in quality. Therefore, the authors claim that more careful
information [28; 10], and HTTP sessions information [107]. and appropriate choice of a machine learning model is very
important.
3.1 Content-basedSpamDetection
Another group introduces features based on HTML page
Seminallineofworkoncontent-basedanti-spamalgorithms structure to detect script-generated spam pages [103]. The
has been done by Fetterly et al. [43; 44; 42; 85]. In [44] underlying idea, that spam pages are machine-generated,
theyproposethatwebspampagescanbeidentifiedthrough is similar to the work discussed above [42; 43]. However,
statistical analysis. Since spam pages are usually automat- here authors make a non-traditional preprocessing step by
ically generated, using phrase stitching and weaving tech- removing all the content and keeping only layout of a page.
niques [54] and aren’t intended for human visitors, they Thus, they study page duplication by analyzing its layout
exhibit anomalous properties. Researchers found that the and not content. They apply fingerprinting techniques [91;
URLsofspampageshaveexceptionalnumberofdots,dashes, 21] with the subsequent clustering to find groups of struc-
digits and length. They report that 80 of the 100 longest turally near-duplicate spam pages.
discovered host names refer to adult websites, while 11 re- There is a line of work dedicated to language modelling for
fertofinancial-credit-relatedwebsites. Theyalsoshowthat spamdetection. [81]presentsanapproachofspamdetection
pages themselves have a duplicating nature – most spam inblogsbycomparingthelanguagemodels[59]forblogcom-
pagesthatresideonthesamehosthaveverylowwordcount ments and pages, linked from these comments. The under-
variance. Another interesting observation is that the spam lyingideaisthatthesemodelsarelikelytobesubstantially
pages’contentchangesveryrapidly9. Specifically,theystud- different for a blog and a spam page due to random nature
ied average amount of week-to-week changes of all the web of spam comments. They use KL-divergence as a measure
pagesonagivenhostandfoundthatthemostvolatilespam of discrepancy between two language models (probability
hostscanbedetectedwith97.2%basedonlyonthisfeature. distributions) Θ , Θ :
1 2
All the proposed features can be found in the paper [44].
Intheirotherwork[42;43]theystudiedcontentduplication (cid:2) p(w|Θ )
KL(Θ (cid:17)Θ )= p(w|Θ )log 1 . (2)
and found that the largest clusters with a duplicate con- 1 2 1 p(w|Θ )
tent are spam. To find such clusters and duplicate content w 2
they apply shingling [22] method based on Rabin finger-
Thebeneficialtraitofthismethodisthatitdoesn’trequire
prints [91; 21]. Specifically, they first fingerprint each of
any training data.
n words on a page using a primitive polynomial PA, sec-
[101; 89] extend the analysis of linguistic features for web
ond they fingerprint each token from the first step with a
spam detection by considering lexical validity, lexical and
different primitive polynomial PB using prefix deletion and
contentdiversity,syntacticaldiversityandentropy,emotive-
extensiontransformations,thirdtheyapplymdifferentfin-
ness, usage of passive and active voices, and various other
gerprinting functions to each string from the second stage
NLP features.
and retain the smallest of the n resulting values for each
of the m fingerprinting functions. Finally, the document is Finally, [10] proposes a number of features based on occur-
representedasabagofmfingerprintsandclusteringisper- renceofkeywordsonapagethatareeitherofhighadvertis-
formedbytakingthetransitiveclosureofthenear-duplicate ing value or highly spammed. Authors investigate discrim-
relationship10. They also mined the list of popular phrases inative power of the following features: Online Commercial
bysorting(i,s,d)triplets11lexicographicallyandtakingsuf- Intention (OCI) value assigned to a URL by Microsoft Ad-
ficiently long runs of triples with matching i and s values. Center12, Yahoo! Mindset classification of a page as either
Based on this study they conclude that starting from the commercialornon-commercial13,GoogleAdWordspopular
36th position one can observe phrases that are evidence of keywords14,andnumberofGoogleAdSenseadsonapage15.
machine-generated content. These phrases can be used as They report an increase in accuracy by 3% over the [24]
an additional input, parallel to common spam words, for a whichdoesn’tusethesefeatures. Similarideaswereapplied
”bag of word”-based spam classifier. for cloaking detection. In [28] search engine query logs and
online advertising click-through logs are analyzed with re-
In [85] they continue their analysis and provide a handful
spect to query popularity and monetizability. Definition of
of other content-based features. Finally, all these features
querypopularityisstraightforward,querymonetizabilityis
are blended in a classification model within C4.5, boosting,
defined as a revenue generated by all ads that are served
and bagging frameworks. They report 86.2% true positive
in response to the query16. Authors use top 5000 queries
and 97.8% true negative rates for a boosting of ten C4.5
out of each query category, request top 200 links for each
trees. Recent work [40] describes a thorough study on how
queryfourtimesbyprovidingvariousagent-fieldstoimitate
various features and machine learning models contribute to
requests by a user (u) and a crawler (c), and then apply a
the quality of a web spam detection algorithm. The au-
cloaking test (3), which is a modified version of a test pro-
thors achieved superior classification results using state-of-
the-artlearningmodels,LogitBoostandRandomForest,and
9it can even change completely on every request. 12adlab.msn.com/OCI/oci.aspx
10two documents are near-duplicates if their shingles agree 13mindset.research.yahoo.com
in two out of six of the non-overlapping runs of fourteen 14adwords.google.com/select/keywordtoolexternal
shingles [42] 15google.com/adsense
11s is the ith shingle of document d 16overlapbetweenthesetwocategoriesisreportedtobe17%
SIGKDD Explorations Volume13,Issue 2 Page53
posed in the earlier work on cloaking detection [108; 110]: • PageRank[86]useslinkinformationtocomputeglobal
importance scores for all pages on the web. The key
min[D(c ,u ),D(c ,u )]
CloakingScore(p)= 1 1 2 2 , (3) underlyingideaisthatalinkfromapagepi toapage
max[D(c ,c ),D(u ,u )]
1 2 1 2 pj shows an endorsement or trust of page pi in page
where pj,andthealgorithmfollowstherepeatedimprovement
|a ∩a | principle, i.e. the true score is computed as a conver-
D(a1,a2)=1−2|a1∪a2| (4) gencepointofaniterativeupdatingprocess. Themost
1 2 popularandsimplewaytointroducePageRankisalin-
isanormalizedtermfrequencydifferencefortwocopiesofa ear system formulation. In this case PageRank vector
pagerepresentedasasetofterms,maybewithrepetitions. for all pages on the web is defined as the solution of
They report 0.75 and 0.985 precision at high recall values the matrix equation
forpopularandmonetizablequeriescorrespondingly,which
(cid:10)π=(1−c)·MT(cid:10)π+c·(cid:10)r, (6)
suggests that the proposed technique is useful for cloaking
detection. However,themethodsdescribedin[28;108;110] wherecisadampingfactor,and(cid:10)risastaticPageRank
might have relatively high false positive rate, since legiti- vector.(cid:8) For non-(cid:9)personalized PageRank it is a unit
mate dynamically generated pages generally contain differ- vector 1,..., 1 , where N =|V|. It is worth noting
N N
ent terms and links on each access, too. To overcome this thatthisisneededtomeettheconditionsofaPerron-
shortcoming an improved method was proposed [75], which Frobeniustheorem[15]andguaranteetheexistenceof
is based on the analysis of structural (tags) properties of astationarydistributionforthecorrespondingMarkov
the page. The idea is to compare multisets of tags and not chain. Ithasalsoasemanticmeaningthatatanygiven
words and links of pages to compute the cloaking score. moment “random surfer” can visit any page with the
non-zero probability.
3.2 Link-basedSpamDetection
Foranon-uniformstaticvector(cid:10)rthesolutioniscalled
Alllink-basedspamdetectionalgorithmscanbesubdivided
apersonalizedPageRank(PPR)[46;56]andthevector
into five groups. The first group exploits the topological
(cid:10)r is called a personalization, random jump or telepor-
relationship (distance, co-citation, similarity) between the
tation vector.
webpagesandasetofpagesforwhichlabelsareknown[86;
55; 71; 99; 112; 12; 52; 26; 66; 11; 5; 8]. The second group There are a few useful properties of PageRank. First,
of algorithms focuses on identification of suspicious nodes itislinearin(cid:10)r,i.e. if(cid:10)π1 isthesolutionof (6)withthe
and links and their subsequent downweighting [16; 84; 73; personalization vector(cid:10)r1 and(cid:10)π2 is the solution of (6)
93; 74; 35; 109; 32; 111; 115; 6]. The third one works by with the personalization vector (cid:10)r2, then the vector
extractinglink-basedfeaturesforeachnodeandusevarious (cid:3)π1+(cid:3)π2 isthesolutionfortheequation(6)withtheper-
2
machine learning algorithms to perform spam detection [4; sonalizationvector (cid:3)r1+(cid:3)r2. Asaconsequenceofthatwe
2
7; 38; 9]. The fourth group of link-based algorithms uses can expand a personalized PageRank in the following
the idea of labels refinement based on web graph topology, way:
whenpreliminarylabelspredictedbythebasealgorithmare (cid:2)
1
modified using propagation through the hyperlink graph or PPR((cid:10)r)= PPR(χv), (7)
N
a stackedclassifier[25; 47; 49; 48]. Finally,thereisa group v∈V
of algorithms which is based on graph regularization tech-
where χv is the teleportation vector consisting of all
niques for web spam detection [121; 1; 30].
zeros except for a node v such that χv(v)=1 (we use
3.2.1 Preliminaries thispropertyinSection3.2.2). Second,PageRankhas
aninterpretationasaprobabilityofarandomwalkter-
minatingatagivenvertexwherethelengthfollowsthe
• Web Graph Model. We model the Web as a graph geometric distribution [62; 45], i.e. the probability to
G =(V,E) with vertices V, representing web pages17, makej stepsbeforeterminationisequaltoc·(1−c)j
anddirectedweightededgesE,representinghyperlinks and the following representation is valid
between pages. If a web page pi has multiple hyper- (cid:2)∞
links to a page pj, we will collapse all these links into PPR((cid:10)r)=c(cid:10)r· (1−c)j(MT)j. (8)
one edge (i,j)∈E. Self loops aren’t allowed. We de- j=0
noteasetofpageslinkedbyapagepi asOut(pi)and
a set of pages pointing to pi as In(pi). Finally, each • HITS[68]algorithmassignshub andauthority scores
edge (i,j) ∈ E can have an associated non-negative for each page on the web and is based on the follow-
weight wij. A common strategy to assign weights is ing observation. Page pi is a good hub, has a high
wij = |Out1(pi)|, though other strategies are possible. hpuagbessc;oarnedhpi,aigfeitispoaingtosotdoamutahnoyrigtoyoidf i(tauisthroefreitraenticveed)
Forinstance,in[2]theyassignweightsproportionalto
by many good hubs, and therefore has a high author-
a number of links between pages. In a matrix nota-
tion a web graph model is represented by a transition ity score ai. As we see the algorithm also has a re-
peatedimprovement principlebehindit. Initsoriginal
matrix M defined as
(cid:17) formthealgorithmconsiderspagesrelevanttoaquery
wij, if(i,j)∈E, based on a keyword-based ranking (the root set), all
Mij = (5)
0, otherwise. the pages that point to them, and all the pages ref-
erenced by these pages. For this subset of the Web
17one can also analyze host level web graph an adjacency matrix is defined, denoted as A. The
SIGKDD Explorations Volume13,Issue 2 Page54
corresponding authority and hub scores for all pages [112] further researches the the idea of propagation by an-
fromthesubsetareformalizedinthefollowingpairof alyzing how trust and distrust propagation strategies can
equations: work together. First, they challenge the way trust is prop-
(cid:17)
agatedinTrustRankalgorithm–eachchild18 getsanequal
(cid:10)a=AT(cid:10)h, (9) part of parent’s trust c · TR(p) , and propose two more
(cid:10)h=A(cid:10)a. |Out(p)|
strategies:
Itcanbeshownthatthesolution((cid:10)h,(cid:10)a)forthesystem • constant splitting, when each child gets the same dis-
of equations (9) after repeatitive updating converges counted part of parent’s trust c·TR(p) score without
to the principal eigenvectors of AAT and ATA corre-
respect to number of children;
spondingly.
• logarithmic splitting, when each child gets an equal
[83] studies the robustness of PageRank and HITS algo-
partofparent’sscorenormalizedbylogarithmofnum-
rithms with respect to small perturbations. Specifically, ber of children c· TR(p) .
they analyzed how severely the ranks will change if a small log(1+|Out(p)|)
portionoftheWebismodified(removed). Theyreportthat
They also analyze various partial trust aggregation strate-
PageRankisstabletosmallperturbationsofagraph,while
gies, whereas TrustRank simply sums up trust values from
HITSisquitesensitive. [18]conductsacomprehensiveanal-
eachparent. Specifically,theyconsidermaximumsharestrat-
ysis of PageRank properties and how link farms can affect
egy, when the maximum value sent by parents is used; and
theranking. Theyprovethatforanylinkfarmandanyset
maximum parent strategy, when propagation is performed
of target pages the sum of PageRank scores over the target
to guarantee that a child score wouldn’t exceed maximum
pagesisatleastalinearfunctionofthenumberofpagesin
ofparentsscores. Finally,theyproposetousealinearcom-
a link farm.
bination of trust and distrust values:
[53] studies optimality properties of link farms. They de-
rivetheboostingfactorforone-targetspamfarmandprove TotalScore(p)=η·TR(p)−β·AntiTR(p), (10)
that thisfarmisoptimalifalltheboosting pagesin a farm
whereη, β ∈(0,1). According totheirexperiments,combi-
have no links between them, they point to the target page
nation of both propagation strategies result in better spam
and target page links back to a subset of them. They also
demotion (80% of spam sites disappear from the top ten
discussmotivationsofspammerstocollude(formalliances),
bucketsincomparisonwiththeTrustRankandPageRank),
and study optimality of web spam rings and spam quasi-
maximum share with logarithmic splitting is the best way
cliques. There is also a relevant work which analyzes prop-
tocomputetrustanddistrustvalues. Theideaoftrustand
ertiesofpersonalizedPageRank[72]andasurveyonefficient
distrust propagation in the context of reputation systems
PageRank computation [14].
was studied in [51].
3.2.2 Algorithmsbasedonlabelspropagation Two algorithms [12; 52] utilize PageRank decomposition
property (Section 3.2.1) to estimate the amount of unde-
The key idea behind algorithms from this group is to con-
served PageRank coming from suspicious nodes. In [12]
sider a subset of pages on the web with known labels and
the SpamRank algorithm is proposed; it finds supporters
then compute labels of other nodes by various propagation
for a page using Monte Carlo simulations [46], assigns a
rules. One of the first algorithms from this category is
penalty score for each page by analyzing whether personal-
TrustRank[55],whichpropagatestrustfromasmallseedset
of good pages via a personalized PageRank. The algorithm izedPageRankscore PPR(χ(cid:10)j)i isdistributed withthebias
towards suspicious nodes, and finally computes SpamRank
relyontheprincipleofapproximateisolation ofagoodset–
for each page as a PPR with the personalization vector ini-
goodpagespointmostlytogoodpages. Toselectaseedset
tialized with penalty scores. The essence of the algorithm
ofreputablepagestheysuggestusinganinversePageRank,
is in the penalty scores assignment. Authors partition all
which operates on a graph with all edges reversed. Having
supporters for a page by their PageRank scores using bin-
computedinversePageRankscoreforallpagesontheWeb,
ning with exponentially increasing width, compute the cor-
theytaketop-Kpagesandlethumanannotatortojudgeon
relation between the index of a bin and the logarithm of a
reputation of these pages. Then they construct a person-
count in the bin, and then assign penalty to supporters by
alization vector where components corresponding only to
summingupcorrelationscoresofpageswhichtheysupport.
reputable judged pages are non-zero. Finally, personalized
The insight behind the proposed approach is that PageR-
PageRankiscomputed. TrustRankshowsbetterproperties
ank follows power law distribution [87]. The concept of
than PageRank for web spam demotion.
spammass wasintroducedin[52]. Spammassmeasuresthe
ThefollowupworkontrustpropagationisAnti-TrustRank[71].
amount of PageRank that comes from spam pages. Similar
Opposite to TrustRank they consider distrust propagation
toTrustRankitneedsthecore ofknowngoodpagestoesti-
from a set of known spam pages on an inverted graph. A
matetheamountofPageRankcomingfromgoodpages. The
seedsetisselectedamongpageswithhighPageRankvalues.
algorithm works in two stages. First, it computes PageR-
They found that their approach outperforms TrustRank on
ank(cid:10)π and TrustRank(cid:10)π(cid:4) vectors and estimates the amount
the task of finding spam pages with high precision and is of spam mass for each page using the formula m(cid:10) = (cid:3)π−(cid:3)π(cid:2).
able to capture spam pages with higher PageRank values (cid:3)π
Second, the threshold decision, which depends on the value
than TrustRank. There is also an algorithm, called Bad-
ofspammass,ismade. Itisworthnotingthatthealgorithm
Rank [99], which proposes the idea to compute badness of
can effectively utilize knowledge about bad pages.
a page using inverse PageRank computation. One can say
that the relation between PageRank and TrustRank is the 18inthisparagraphwewillrefertoout-neighboursofapage
same as between BadRank and Anti-TrustRank. as children and in-neighbours as parents
SIGKDD Explorations Volume13,Issue 2 Page55
Credibility-based link analysis is described in [26]. In this andprovethetheoremthatanydampingfunctionsuchthat
work the authors define the concept of k-Scoped Credibility the sum of dampings is 1 yields a well-defined normalized
foreachpage,proposeseveralmethodsofitsestimation,and functional ranking. They study exponential (PageRank),
show how it can be used for web spam detection. Specifi- linear,quadratichyperbolic(TotalRank),generalhyperbolic
cally,theyfirstdefinetheconceptofBadPath,ak-hopran- (HyperRank)dampingfunctions,andproposeefficientmeth-
domwalkstartingfromacurrentpageandendingataspam ods of rank computation. In [8] they research the appli-
page,andthencomputethetunedk-ScopedCredibility score cation of general damping functions for web spam detec-
as tionandproposetruncatedPageRank algorithm,whichuses
(cid:18) (cid:2)k (cid:19) (cid:2) (cid:8) (cid:9)(cid:20)(cid:21) truncated exponential model. The key underlying observa-
Ck(p)= 1− P pathl(p) ·γ(p), tion behind the algorithm is that spam pages have a large
number of distinct supporters at short distances, while this
l=1 pathl(p)∈BadPathl(p)
(11) number is lower than expected at higher distances. There-
where k is a parameter specifying the length of a random fore, they suggest using damping function that ignore the
walk, γ(p) is a credibility penalty factor that is needed to direct contribution of the first few levels of in-links
(cid:17)
deal with only(cid:8) partial(cid:9)know(cid:22)ledge of all spam pages on the 0 if j ≤J,
Web19,andP pathl(p) = li−=10wii+1. Thecredibilityscore damping(j)= D(1−c)j otherwise. (14)
canbeusedtodownweightorprunelowcrediblelinksbefore
link-based ranking or to change the personalization vector In this work they also propose a probabilistic counting al-
in PPR, TrustRank, or Anti-TrustRank. gorithm to efficiently estimate number of supporters for a
In [66] the concept of anchor is defined, as a subset of page. Link-based feature analysis and classification mod-
pageswithknownlabels,andvariousanchor-basedproxim- els using link-based and content-based features are studied
ity measures on graphs are studied. They discuss personal- in [7; 9] correspondingly.
izedPageRank;harmonicrank,whichisdefinedviarandom
walk on a modified graph with an added source and a sink 3.2.3 Linkpruningandreweightingalgorithms
such that all anchor vertices are connected to a source and Algorithms belonging to this category tend to find unreli-
all vertices are connected to a sink with probability c; non- able links and demote them. The seminal work [16] raises
conserving rank, which is a generalization of personalized problemsinHITS[68]algorithm,suchasdominationofmu-
PageRank satisfying the equation tually reinforcing relationships and neighbour graph topic
(cid:10)π=(I−(1−c)·MT)−1(cid:10)r. (12) drift, and proposed methods of their solution by augment-
ing a link analysis with a content analysis. They propose
They report that non-conserving rank is the best for trust to assign each edge an authority weight of 1 if there are k
k
propagation, while harmonic rank better suits for distrust pagesfromonesitelinktoasinglepageonanothersite,and
propagation. assign a hub weight of 1 if a single page from the first site
l
Spam detection algorithm utilizing pages similarity is pro- is pointing to l pages on the other site. To combat against
posedin[11],wheresimilarity-basedtop-Klistsareusedto topic drift they suggest using query expansion, by taking
compute a spam score for a new page. Authors consider top-Kfrequentwordsfromeachinitiallyretrievedpage;and
co-citation,CompanionRank,SimRank[61],andkNN-SVD candidate page set pruning, by taking page relevance as a
projectionsasmethodstocomputesimilaritybetweenpages. factor in HITS computation. The same problems are stud-
First, for a page to be classified a top-K result list is re- iedin[84],whereaprojection-basedmethodisproposedto
trieved using some similarity measure. Second, using the compute authority scores. They modify eigenvector part of
retrievedpagesthefollowingfourvaluesarecomputed: frac- HITSalgorithminthefollowingway. Insteadofcomputing
tion of the number of labeled spam pages in the list (SR), a principal eigenvector of ATA, they compute all eigenvec-
anumberoflabeledspampagesdividedbyanumberofla- tors of the matrix and then take the eigenvector with the
beled good pages in the list (SON), sum of the similarity biggestprojectionontherootset(setofpagesoriginallyre-
valuesoflabeledspampagesdividedbythetotalsimilarity trieved by keyword search engine, as in HITS), finally they
value of pages retrieved (SVR), and the sum of the simi- reportauthorityscoresasthecorrespondingcomponentsof
larity values of labeled spam pages divided by the sum of this eigenvector.
thesimilarityvaluesoflabeledgoodpages(SVONV).Third, Another group introduces the concept of tightly-knit com-
threshold-based rule is used to make a decision. According munity (TKC) and proposes SALSA algorithm [73], which
totheirexperiments,similarity-basedspamdetection(SVR, performs two random walks to estimate authority and hub
SVONV)performsbetterathighlevelsofrecall,whileAnti- scoresforpagesinasubgraphinitiallyretrievedbykeyword-
TrustRank[71]andcombinedTrust-Distrust[112]algorithms basedsearch. Itisworthnotingthattheoriginalandanin-
show higher precision at low recall levels. verted subgraphs are considered to get two different scores.
TheseminallineofworkwasdonebyBaeza-Yatesetal.[5; Anextensionofthiswork[93]considersclusteringstructure
8; 9; 7]. In [5], inspired by the PageRank representation onpagesandtheirlinkagepatternstodownweightbadlinks.
(Equation 8), they propose the concept of functional rank, The key trick is to count number of clusters pointing to a
which is a generalization of PageRank via various damping page instead of number of individual nodes. In this case
functions. Theyconsiderrankingbasedonageneralformula authority of a page is defined as follows:
(cid:2)
p(cid:10)= N1(cid:10)1(cid:2)∞ damping(j)(MT)j, (13) aj =k:j∈l(k)(cid:10)i:j∈1l(i)Sik, (15)
j=0
19several strategies to define γ are proposed. where Sik = ||ll((ii))∩∪ll((kk))|| and l(i) is a set of pages linked from
SIGKDD Explorations Volume13,Issue 2 Page56
pagepi. Theapproachactslikepopularityrankingmethods ysis. [4]studieslink-basedfeaturestoperformwebsitecate-
discussed in [20; 27]. gorizationbasedontheirfunctionality. Theirassumptionis
[74] studies “small-in-large-out” problem of HITS and pro- thatsitessharingsimilarstructuralpatterns,suchasaverage
posestoreweightincomingandoutgoinglinksforpagesfrom page level or number of outlinks per leaf page, share simi-
the root set in the following way. If there is a page whose lar roles on the Web. For example, web directories mostly
in-degree is among the three smallest ones and whose out- consist of pages with high ratio of outlinks to inlinks, form
degreeisamongthethreelargestones,thensettheweight4 a tree-like structure, and the number of outlinks increases
forallin-linksofallrootpages,otherwisesetto1. Runone with the depth of a page; while spam sites have specific
iteration of HITS algorithm without normalization. Then topologies aimed to optimize PageRank boost and demon-
if there exists a root page whose authority value is among strate high content duplication. Overall, each website is
the three smallest ones and whose hub value is among the represented as a vector of 16 connectivity features and a
threelargestones,settheweight4forallin-linksofallroot clusteringisperformedusingcosineasasimilaritymeasure.
pages, and then run the HITS algorithm again. Authorsreportthattheymanagedtoidentify183webspam
[35]introducestheconceptof“neponistic”links–linksthat rings forming 31 cluster in a dataset of 1100 sites.
present for reasons rather than merit, for instance, naviga- Numerouslink-basedfeatures,derivedusingPageRank,Trust-
tional links on a website or links between pages in a link Rank, and truncated PageRank computation are studied
farm. They apply C4.5 algorithm to recognize neponis- in [7]. Mixture of content-based and link-based features
tic links using 75 different binary features such as IsSim- is used to combat against web spam in [38; 9], spam in
ilarHeaders, IsSimilarHost, is number of shared in-links is blogs [69; 76].
greater than a threshold. Then they suggest pruning or
3.2.5 Algorithmsbasedonlabelsrefinement
downweighting neponistic links. In their other work [109]
theycontinuestudyinglinksindenselyconnectedlinkfarms. The idea of labels refinement has been studied in machine
Thealgorithmoperatesinthreestages. First,itselectsaset learning literature for general classification problems for a
of bad seed pages guiding by the definition that a page is long time. In this section we present algorithms that ap-
bad if intersection of its incoming and outgoing neighbours ply this idea for web spam detection. In [25] a few web
isgreaterthanathreshold. Second,itexpandsthesetofbad graph-based refinement strategies are proposed. The algo-
pages following the idea that a page is bad if it points to a rithm in [25] works in two stages. First, labels are assigned
lotofbadpagesfromtheseedset. Finally,linksbetweenex- using a spam detection algorithm discussed in [7]. Then,
panded set of bad pages are removed or downweighted and at the second stage labels are refined in one of three ways.
any link-based ranking algorithm [86; 68] can be applied. One strategy is to perform Web graph clustering [67] and
Similar ideas are studied on a host level in [32]. then refine labels guided by the following rules. If the ma-
In [111] the concept of a complete hyperlink is proposed, a jority of pages in a cluster is predicted to be spam, they
hyperlink coupled with the associated anchor text, which denote all pages in the cluster as spam. Formally, they as-
is used to identify pages with suspiciously similar linkage sume that predictions of the base algorithm are in [0,1],
patterns. Rationalebehindtheirapproachisthatpagesthat then they compute the average value over the cluster and
have high complete hyperlink overlap are more likely to be compareitagainstathreshold. Thesameprocedureisdone
machine-generated pages from a link farm or pages with fornon-spamprediction. Theotherstrategyoflabelrefine-
duplicating content. The algorithm works as follows. First, ment,whichisbasedonpropagationwithrandomwalks,is
itbuildsabasesetofdocuments,asinHITS,andgenerates proposed in [120]. The key part is to initialize the person-
a page-hyperlink matrix using complete hyperlinks, where alization vector(cid:10)r in PPR by normalizing the predictions of
Aij =1,ifapagepicontainscomplete-hyperlinkchlj. Then the base algorithm: rp = (cid:2)p∈s(Vp)s(p), where s(p) is a predic-
it finds bipartite components with the size greater than a tion of the base algorithm and rp is the component of the
thresholdinthecorrespondinggraph,wherepartsarepages vector(cid:10)rcorrespondingtopagep. Finally,thethirdstrategy
and links, and downweight complete hyperlinks from large is to use stacked graphical learning [70]. The idea is to ex-
components. Finally, a HITS-like algorithm is applied on a tend the original feature representation of an object with a
reweighted adjacency matrix. newfeaturewhichisanaveragepredictionforrelatedpages
[115] notices that PageRank score of pages that achieved in the graph and run a machine learning algorithm again.
highranksbylink-spammingtechniquescorrelateswiththe They report 3% increase over the baseline after two rounds
damping factor c. Using this observation authors identify of stacked learning.
suspiciousnodes,whosecorrelationishigherthanathresh- Afewotherrelabellingstrategiesareproposedin[47;49;48].
old, and downweight outgoing links for them with some [47] suggests constructing an absolutely new feature space
function proportional to correlation. They also prove that by utilizing predictions from the first stage: the label by
spammers can amplify PageRank score by at most 1c and the base classifier, the percentage of incoming links coming
experimentallyshowthateventwo-nodecollusioncanyield fromspamandpercentageoutgoinglinkspointingtospam,
a big PageRank amplification. The follow-up work [6] per- etc. Overall,theyconsidersevennewfeaturesandreportin-
formsmoregeneralanalysisofdifferentcollusiontopologies, creaseinperformanceoverthebaseclassifier. [48]proposes
where they show that due to the power law distribution of tousefeaturere-extractionstrategyusingclustering,propa-
PageRank [87], the increase in PageRank is negligible for gation, and neighbour-graph analysis. Self-training was ap-
top-ranked pages. The work is similar to [53; 3]. plied in [49] so as to reduce the size of the training dataset
requiring for web spam detection.
3.2.4 Algorithmswithlink-basedfeatures
3.2.6 Graphregularizationalgorithms
Algorithms from this category represent pages as feature
vectorsandperformstandardclassificationorclusteringanal- Algorithmsinthisgroupperformtrunsductiveinferenceand
SIGKDD Explorations Volume13,Issue 2 Page57
utilize Web graph to smooth predicted labels. According URLinreply(weight2),viewURLinpreviousposts(weight
toexperimentalresults,graphregularizationalgorithmsfor 1). Then they build a user-URL bipartite graph where an
web spam detection can be considered as the state-of-the- edge weight is the sum of all weights associated with the
art at the time of writing. The work [121] builds a discrete corresponding actions. After that they compute SimRank
analogue of classification regularization theory [102; 104] for all pairs of URLs and introduce the regularization term
by defining discrete operators of gradient, divergence and via Laplacian of the similarity matrix. The second regular-
Laplacianondirectedgraphsandproposesthefollowingal- ization term is defined analogously, but now simply via a
gorithm. First,theycomputeaninverseweightedPageRank Laplacianofastandardtransitionmatrix(seeeq.5). Third
withtransitionprobabilitiesdefinedasaij = Inw(jpii). Second, and fourth asymmetric terms, defined via the Web graph
they build the graph Laplacian transitionmatrixanditsdiagonalroworcolumnaggregated
matrices, are introduced to take into account the principle
ΠA+ATΠ
L=Π−α , (16) of approximate isolation of good pages. Finally, the sound
2 quadratic problem is solved. Authors report that even le-
whereαisauser-specifiedparameterin[0,1],Aisatransi- gitimately looked spam websites can be effectively detected
tionmatrix,andΠisadiagonalmatrixwiththePageRank usingthismethodandhencetheapproachcomplementsthe
score20 over the diagonal. Then, they solve the following existing methods.
matrix equation
3.3 Miscellaneous
Lϕ(cid:10) =Π(cid:10)y, (17)
Inthissectionwediscussalgorithmsthatusenon-traditional
where(cid:10)yisavectorconsistingof{−1,0,1}suchthatyi =1if features and ideas to combat against web spam.
pageisnormal,yi =−1ifitisspam,andyi =0ifthelabel
3.3.1 Unsupervisedspamdetection
for a page is unknown. Finally, the classification decision
is made based on the sign of the corresponding component Theproblemofunsupervisedwebspamdetectionisstudied
of vector ϕ(cid:10). It is worth noting that the algorithm requires in[119;118;116]. Authorsproposetheconceptofspamicity
strongly connected graphs. anddevelopanonlineclient-sidealgorithmforwebspamde-
Another algorithm that follows regularization theory is de- tection. The key distinctive feature of their solution is that
scribed in [1]. There are two principles behind it. First, itdoesn’tneedtrainingdata. Atthecoreoftheirapproach
it addresses the fact that hyperlinks are not placed at ran- isa(θ,k)-pagefarmmodel introducedin[117],thatallowsto
dom implies some degree of similarity between the linking computethetheoreticalboundonthePageRankscorethat
pages [36; 27] This, in turn, motivates to add a regularizer can be achieved using the optimal page farm with a given
to the objective function to smooth predictions. Second, numberofpagesandlinksbetweenthem. Theproposedal-
it uses the principle of approximate isolation of good pages gorithmworksasfollows. Foragivenpageitgreedilyselect
that argues for asymmetric regularizer. The final objective pages from the k-neighbourhood, that contribute most to
function has the following form: the PageRank score, following the definition
(cid:17)
(cid:2)l PR[p,G]−PR[p,G(V−{v})], ifv(cid:5)=p,
Ω(w(cid:10),(cid:10)z)= 1l L(w(cid:10)T(cid:10)xi+zi,yi)+λ1(cid:17)w(cid:10)(cid:17)2+λ1(cid:17)(cid:10)z(cid:17)2 PRContrib(v,p)= 1N−c, otherwise,
i=1 (cid:2) (19)
+γ aijΦ(w(cid:10)T(cid:10)xi+zi,w(cid:10)T(cid:10)xj+zj), (18) and at each iteration it computes the link-spamicity score
(i,j)∈E as the ratio of the observed PageRank contribution from
selectedpagesovertheoptimalPageRankcontribution. The
where L(a,b) is a standard loss function, w(cid:10) is a vector of
algorithm uses monotonicity property to limit number of
coefficients, (cid:10)xi and yi are feature representation and a true supportersthatneedstobeconsidered. Finally,itmarksthe
label correspondingly, zi is a bias term, aij is a weight of pageassuspiciousiftheentirek-neighbourhoodofthispage
the link (i,j) ∈ E, and Φ(a,b) = max[0,b−a]2 is the reg-
is processed and link-spamicity score is still greater than a
ularization function. The authors provide two methods to
threshold. Analogous optimality conditions were proposed
find the solution of the optimization problem using conju-
for a page content. In this case term-spamicity score of a
gategradientandalternatingoptimization. Theyalsostudy
pageisdefinedastheratiooftheTFIDFscoreachievedby
theissueofweightssettingforahostgraphandreportthat
the observed content of a page over the TFIDF score that
thelogarithmofthenumberoflinksyieldsthebestresults.
could be achieved by an “optimal” page that has the same
Finally, according to the experimental study, the algorithm
number of words.
has good scalability properties.
SomeinterestingideatoextractwebspamURLsfromSEO 3.3.2 Algorithmsbasedonuserbrowsingbehaviour
forums is proposed in [30]. The key observation is that on
[77] proposes to incorporate user browsing data for web
SEO forums spammers share links to their websites to find
spam detection. The idea is to build a browsing graph G =
partners for building global link farms. Researchers pro-
(V,E,T,σ)wherenodesV arepagesandedgesE aretransi-
pose to solve a URL classification problem using features
tionsbetweenpages,T correspondstoastayingtimeandσ
extracted from SEO forum, from Web graph, and from a
denotestherandomjumpprobability,andthencomputean
website, and regularizing it with four terms derived from
importance for each page using PageRank-like algorithm.
link graph and user-URL graph. The first term is defined
The distinctive feature of the proposed solution is that it
as follows. Authors categorize actions on a SEO forum in
considers continuous-time Markov process as an underly-
threegroups: postURLinarootofathread (weight3),post
ingmodelbecauseuserbrowsingdataincludetimeinforma-
20computed at the first stage. tion. Formally, the algorithm works as follows. First, using
SIGKDD Explorations Volume13,Issue 2 Page58
staying times Z1,...,Zmi for a page i it finds the diagonal The idea to use rank-time features in addition to query-
element qii of the matrix Q = P(cid:4)(t) as a solution of the independentfeaturesisintroducedin[100]. Specifically,au-
optimization problem: thors solve the problem of spam pages demotion after the
(cid:23) 1 1 1 (cid:24)2 querywasissuedguidedbytheprinciplethatspammersfool
(Z¯+ qii)− 2(S2− qi2i) −→mqiiin, s.t. qii <0. (20) rmaentkhiondgsarlgaothrietrhtmhsanangdenauchinieevlyerheilgehvapnotsiptaiogness,uasnindgtdhieffreerfeonret
Non-diagonal elements of the matrix Q are estimated as spam pages should be outliers. Overall, 344 rank-time fea-
(cid:17) tures are used such as number of query terms in title and
−qqiiji = cσ(cid:2)j,Nk=w+˜1i1jw˜ij +(1−c)σji,=iN∈+V1,,jj∈∈V˜V, (21) fcaronendqtuaaeinnpcaaygqeou,fefaroyrqtudeeirrffmye,rtenenr-tmgrvaoamnluoeasvpeoraflgaenp,snabnuedmtwbfoeerernodfqiffupeearrgeynesttetsrhkmaipst
whereadditionalN+1pseudo-vertexisaddedtothegraph n-grams. Accordingtotheexperiment,theadditionofrank-
to model the teleportation effect such that all last pages in timefeaturesallowstoincreaseprecisionby25%atthesame
the corresponding browsing sessions are connected to it via levels of recall. In the same work they study the problem
edgeswiththeweightsequaltothenumberofclicksonthe ofoverfittinginwebspamdetectionandsuggestthattrain-
lastpage,andpseudo-vertexisconnectedtothefirstpagein ingandtestingdatashouldbedomain-separated,otherwise
each session with the weight equal to normalized frequency testing error could be up to 40% smaller than the real.
of visits for this page; σj is a random jump probability.
Atthesecondstep,thestationarydistributioniscomputed 3.3.4 Clickspamdetection
using the matrix Q, which is proven to correspond to the Sinceclickspamaimstopush“maliciousnoise”intoaquery
stationarydistributionforthecontinuous-timeMarkovpro- logwiththeintentiontocorruptdata,usedfortheranking
cess. Similar idea is studied in [90]. The authors define a function construction, most of the counter methods study
hyperlink-click graph,whichisaunionofthestandardWeb the ways to make learning algorithms robust to this noise.
graph and a query-page click-log-based, and apply random Otheranti-click-fraudmethodsaredrivenbytheanalysisof
walks to perform pages ranking. the economic factors underlying the spammers ecosystem.
User behaviour data is also used in [78]. There are two Interesting idea to prevent click spam is proposed in [92].
observations behind their approach. First, spammers aim The author suggests using personalized ranking functions,
to get high ranking in SERP and therefore the major part as being more robust, to prevent click fraud manipulation.
of the traffic to spam website is due to visits from a search The paper presents a utility-based framework allowing to
engine site. Second, users can easily recognize a spam site judgewhenitiseconomicallyreasonabletohirespammersto
and therefore should leave it quite fast. Based on these promoteawebsite,andperformsexperimentalstudydemon-
observations authors propose 3 new features: ratio of visits strating that personalized ranking is resistant to spammers
fromasearchsite,numberofclicksandpageviewsonasite manipulationsanddiminishesfinancialincentivesofsiteown-
per visit. ers to hire spammers. The work [37] studies the robustness
of the standard click-through-based ranking function con-
3.3.3 HTTPanalysisandreal-timespamdetection
struction process and also reports its resistance to fraudu-
Methodsinthissubsectioncanbenaturallypartitionedinto lent clicks.
2groups: client-sideandserver-side. Themethodsfromthe The work [60] studies the problem of click fraud for online
first group useverylimited information, usuallydoesn’t re- advertisingplatformandparticularlyaddressestheproblem
quirelearning,andlessaccurate. Thelattergrouprepresen- of “competitor bankruptcy”. The authors present a click-
tative methods, in reverse, are more precise since they can based family of ads pricing models and theoretically prove
incorporate additional real-time information. thatsuchmodelsleavenoeconomicincentivesforspammers
Lightweight client-side web spam detection method is pro- to perform malicious activity, i.e. short term competitor’s
posedin[107]. Insteadofanalyzingcontent-basedandlink- budgetwastingwillbeannihilatedbylongtermdecreasein
basedfeaturesforapage,researchersfocusedonHTTPses- the ads placement price. [105] carefully analyzes the entire
sion information and achieved competitive results. They spammers’ ecosystem by proposing the spam double fun-
represent each page and a session, associated with it, as a nel model which describes the interaction between spam-
vectoroffeaturessuchasIP-addressorwordsfromarequest publishers and advertisers via page redirections.
headerina“bag-of-words”model,andperformaclassifica- In[17]anincentivebasedrankingmodelisintroduced,which
tion using various machine learning algorithms. The same mainlyincorporatesusersintorankingconstructionandpro-
group introduced the way of large dataset creation for web videsaprofitablearbitrageopportunityfortheuserstocor-
spam detection by extracting URLs from email spam mes- rect inaccuracies in the system. The key idea is that users
sages [106]. Though not absolutely clean, the dataset con- are subject to an explicit information about revenue they
tains about 350000 web spam pages. might “earn” within the system if they correct an erro-
Similarly,HTTPsessionsareanalyzed,amongothers,inthe neous ranking. It is theoretically proven that the model
contextofmaliciousredirectiondetectionproblem. Theau- with the specific incentives (revenue) structure guarantees
thorsof[29]provideacomprehensivestudyoftheproblem, merit-based ranking and is resistant to spam.
classifyallspamredirectiontechniquesinto3types(HTTP
redirection,METARefresh,JavaScriptredirection),analyze
4. KEYPRINCIPLES
thedistributionofvariousredirectiontypesontheWeb,and
present a lightweight method to detect JavaScript redirec- Havinganalyzedalltherelatedworkdevotedtothetopicof
tion, which is the most prevalent and difficult to identify webspammining,weidentifyasetofunderlyingprinciples
type. that are frequently used for algorithms construction.
SIGKDD Explorations Volume13,Issue 2 Page59