DSPIN: Detecting Automatically Spun Content on the Web - PowerPoint Presentation

Slide1

DSPIN: Detecting Automatically Spun Content on the Web

Qing Zhang, David Y. Wang, Geoffrey M. VoelkerUniversity of California, San Diego

1Slide2

What is Spinning?

A Black H

at

S

earch Engine Optimization (BHSEO) technique that rewords original content to avoid duplicate detectionTypically an article (seed) is spun multiple times creating N versions of the article that will be posted on N different sitesArtificially generate interest to increase search result rankings of targeted site

2Slide3

Spinning Example

3Slide4

Spinning Approaches

Human Spinning

Hire a

real person

from an online marketplace (i.e. Fiverr, Freelancer) to spin manuallyPros: Reasonable text readabilityCons: Expensive ($2-8 / hr)Not scalable (humans)Automated SpinningRun software to spin automaticallyPros: Fast

Cheap ($5)

Scalable (500 articles / job)

Minimal human interaction

Cons:

Can read awkwardly

4Slide5

Spinning in BHSEO

5

SEO Software

Start with a seed article

and SEO SoftwareSlide6

Spinning in BHSEO

6

SEO Software

SEO Software submits the

a

rticle to spinner (TBS)Slide7

Spinning in BHSEO

7

SEO Software

TBS spins the article and

verifies plagiarism

d

etection failsSlide8

Spinning in BHSEO

8

SEO Software

SEO Software receives

s

pun articleSlide9

http://<

moneysite>

http://<

moneysite

>Spinning in BHSEO9

SEO Software

Proxies

User Generated Content

SEO Software posts articles

on User Generated Content

t

hrough proxiesSlide10

Spinning in BHSEO

10

SEO Software

Proxies

User Generated Content

Search Engine

Search Engine consumes

u

ser generated contentSlide11

Goals

Understand the current state of automated spinning software using one of the most popular spinners

(The Best Spinner)

Develop techniques to detect spinning using immutables + mutablesExamine spinning on the Web using Dspin, our system to identify automatically spun content11Slide12

The B

est Spinner (TBS)TBS consists of

two parts

Program (binary):

provides the user interfaceSynonym dictionary: a homemade, curated list of synonyms that are updated weeklyReplaces text with synonyms from dictionaryWe extract the synonym dictionary through reverse engineering the binary12Slide13

TBS Example

13Slide14

Immutables +

MutablesAn article is composed of

immutables

(NOT IN dictionary) and mutables (IN dictionary)14Slide15

Spinning Detection Algorithm

Immutables detection computes the

ratio of shared

immutables

between two pagesWorks well in practice except in corner case where there are few immutables to compareMutables detection computes the ratio of all shared words after two levels of recursively expanding synonymsAlso works well and handles corner case, but expensive15Slide16

Other Approaches

Duplicate content detection is a well known problem for Search EnginesExplored

other approaches

:

Hashes of substrings [Shingling]Parts of speech [Natural Language Processing]Spinning is designed to circumvent these approaches (i.e. replace every Nth word, synonym phrases)16Slide17

Validation

Setup controlled experiment using TBS600 article test data

set

Started with 30 seed articles

5 articles from 5 different article directories5 articles randomly chosen from Google NewsEach article spun 20 times w/ bulk spin optionImmutables detects all spun content and matches with the source17Slide18

DSpin

Detection from Search Engine POVInput:

set of

article pages

crawled from the WebOutput: set of pages flagged as auto spunBuild graph of clusters of “similar” pages using immutables + mutables approachEach page represents a nodeCreate edges between pairs of nodes using immutables, verify edges using mutablesEach connected components is cluster18Slide19

Results

Ran DSpin on a real life data set

Set of 797 abused wikis

Crawl each wiki

daily for newly posted articlesCollected 1.23M Articles from Dec 2012Address the following questions:Is spinning a problem in the wild?Can we characterize spinning behavior?19Slide20

Filtering

20

Filter out pages that are: non-English, exact duplicates, < 50 words, or primarily

links

225K spun pages remaining.Spinning is for real.Slide21

Wiki Content

21

Spinning campaigns target business + marketing termsSlide22

Cluster Size

12.7K clusters from 225K

spun pages

22

Moderate clusters of spun articles in abused wikisSlide23

Timing Duration

23Duration reveals how long a campaign lasts

Compute by extracting dates, max – min

Most campaigns occur in bursts. Slide24

Conclusion

Proposed + evaluated a spinning detection algorithm based on immutables

+

mutables

that Search Engines can implementDemonstrated the algorithm's applicability on a real life data set (abused wikis)Characterized the behavior of at least one slice of the Web where spun articles thrive24Slide25

Thank You!

Q&A25Slide26

TBS Coverage

Only one synonym dictionary was used to implement DSpin, is this system still applicable widely (i.e. for other spinners)?We had

no prior knowledge

about

how articles from abused wikis were spunYet we still detected spun articles26Slide27

Synonym Dictionary Churn

How much does the synonym dictionary change over time?We re-fetched synonym dictionary four months after the initial study and found that

94% of terms remain the same

Furthermore,

DSpin detected spun articles posted months prior27Slide28

Synonyms in the Cloud

What if the spinner stores the synonym dictionary in the cloud?There is an operational cost for the spinner

(network bandwidth == $$$)

Can

still reconstruct synonym dictionary through controlled experiments (i.e. submitting our own articles for spinning)28Slide29

Scalability

How can Search Engines implement the immutables algorithm?Assume

Search Engines

already perform duplicate content detectionCan think of immutables approach as performing duplicate content detection on the immutables portion of the pages (a subset of what is already currently done)29