A Ranking Algorithm for Semantic search engine – spam and fake detection case study By: Soheila Dehghanzadeh. Web technology lab weekly seminars.

A Ranking Algorithm for Semantic search engine –

spam and fake detection case studyBy: Soheila Dehghanzadeh.Web technology lab weekly

seminars.

Agenda :

• Web spam definition• A brief overview of Search engines• Search engine phases:– Crawling– Indexing– Index lookup– Ranking lookup results

• My proposed ranking algorithm

Web spam and fake:

• In web of data anyone is able to say anything about anything.

• Low quality data should not be mentioned in top search results.

A Search Engine:

A Search Engine:

web of data vs. web of documents.

• WODocNo: link type and no trustworthiness (just popularity).

• WOData: should consider link type and link context (for provenance and proof of trust).

Crawling & Indexing phase…

• Using ldspider to crawl linked data.• Using hexastore for complete indexing the

crawled data. Special thanks to Panagiotis Karras for providing hexastore implementation in python.

•Index lookup results for extension…some Results may not include keyword but they have high quality and relevance.Result expansion to hide the locality effect. Some sites is referred many times but in this special context other professional sites lookup results are more interested.

Web of

data

result

HexaStore

• Index structure that we use in our search engine.• Each RDF element type deserve to have special

index structure build round it. • Every possible ordering of the importance or

precedence of the three elements in the indexing scheme is materialized.

• Each index structure in a hexastore centers around one RDF element and defines a perioritation between the other 2 elements.

Sample spo indexing in a hexastore

Si

P(I,1) P(I,2) P(I,Ni)

O(i1,1)

O(i1,2)

O(i1,ki1)

O(i2,1)

O(i2,2)

O(i2,ki2)

O(iNi,1)

O(iNi,2)

O(iNi,kiNi)

Space complexity :Spo+sp+pso+so+pos+po

My idea!• Import the base result set to jena and extend it.• Extending the base set with ontology reasoning rules so that extra resources and

relations will be added through reasoning rules.• The added resources• The added relation has no context so their trustworthiness is an aggregation

function on (x,y,rule) relations---• Resources will be added only through sameAs predicate • Resources will be ranked according to relevance to query terms (using ontobroker

– pagerank – objectrank- triplerank – HITS,….)• Query

– Keyword query– Structured query– Ontology based query (using an interface to get query) - ontobroker

• Relation (properties) will be ranked according to contexts(provenance) using relation ranking methods such as semRank or we can look at context’s pageRank.

• Note that First we rank resources and second we rank relations . However it depends on the user query whether it is looking for relations or resources.

• Lookup on quads for keyword (Soheila)• Q1: http://um.s11,givenname,”Soheila”,UM• Q2:http://NIOC/p25,fullname,”Soheila Dehghan”,NIOC• Q3:http://nigc-khrz/e66,firstname,”Soheila”,NIGC• Q4: http://fake/f4,name,”Soheila”,fake

Q1http://

facebook/u122http://isport

/us122SA(UM) SA(FB)

Q3SA(UM)

Q2SA(NIGC)

http://linkedIn/u12

SA(LI)

Q4SA(FK)

SA(FK) SA(FK)

CheeseB.Gates

ScottDancewith(FK)

Meet(CNN)

Buy(Spam)

Result set expansion methods:

• step1: using sameas predicate on found Qaudes and extend ResultSet to Q1,…,Qr

• index LookUp– Q1(S),SameAs,?,?Qr(S),SameAs,?,?– ?,SameAs,Q1(S),? ?,SameAs,Qr(S),?

• (Q1,…,Q4Q1,…,Q4,FBURI,LinkedInURI,isportURI) in our case.• apply PR on Extended graph with SameAs which

SameAs links are replaced with PR weight of sameAs context.(to know the trustwothiness of each contexts).

Result set expansion methods:• Step2: LookUp all properties of Q1(s),…,Qr(s)– Q1(s),?,?,?—?,?,Q1(s),?– …– Qr(s),?,?,?—?,?,Qr(s),?

• Step4: add inferred relation using domain ontology(context is composed of ontology+inference process)• Step4: rank Q1,…,Qr according to their TpageRank

(computed online from graph of step1 ), rank relations according to their context pageRank(which is computed by Google offline)

• Note : contexts who has PR lower than a treshhold won’t be mentioned.they maybe Spam or Fake Sites.

Structured query on quads indexes

• Single pivot:– (S,?,?,?),(?,p,?,?),(?,?,o.?),(?,?,?.C)

• Double pivot:– (S,p,?,?),(s,?,O,?),(s,?,?,C),(?,P,O,?),(?,P,?,C),

(?,?,o,C)

• Triple pivot:– (s,p,o,?),(s,p,?,c),(s,?,o,c),(?,p,o,c)

• Step1: if the specified parts was URI then a direct lookup is performed by search engine . Otherwise if user have specified keyword for each parts then firstly a keyword search will be done and then for each result URI a lookup will be performed.

Lookup on quads for ontological queries

Soheila(FUM)

FUM

GAS

NIOC Team

Sally(NIOC)

Studied in(FUM)

Worked at(GAS)

played in(NIOC)

Owl:sameAs(NIOC)

DehghanZadeh(GAS)

Owl:sameAs(FUM)

Kahani(FUM)

Supervisor(FUM)

Related works for ranking web of data…

• Objectrank• Ding• Sindice ti-idf• EntityRank.• semRank• ReConRank• ontobroker…

Proof of trust

• Jena inference Explanation will be used to represent as a proof of trust

Evaluation …

• Compare Spam ranks • Compare query time• Compare index size

Any question?

Best things in the life are

free.Thanks for attention.