Follow the Green: Growth and Dynamics on Twitter Follower Markets

Follow the Green: Growth and Dynamics

in Twitter Follower Markets

Gianluca Stringhini, Gang Wang, Manuel Egele*, Christopher Kruegel, Giovanni Vigna, Ben Y. Zhao, Haitao Zheng

UC Santa Barbara *Carnegie Mellon University

Twitter Followers = Perceived Reputation

Gianluca Stringhini – Follow the Green: Growth and Dynamics in Twitter Follower Markets

Building a network of followers is difficult!

2

Services that measure the Twitter influence of an account (such as Klout) take the number of followers into account, together with a number of other indicators

Shortcuts to Success

Gianluca Stringhini – Follow the Green: Growth and Dynamics in Twitter Follower Markets 3

Can One Really Buy Followers?


Twitter Follower Markets

• Fake accounts (Sybils)• Compromised accounts• Pyramid schemes


Different types of followers for sale

5

Pyramid Markets


• Free subscribers → Victims• Paid subscribers → Customers

Twitter’s ToS forbids users to participate in Twitter Follower Markets

6

Free Subscriber

Paid Subscriber

Our Contributions

• We study the Twitter Follower Market phenomenon• We analyze the characteristics of market customers and victims• We can detect accounts that bought followers

Twitter could block such accounts Twitter Follower Markets would go bankrupt


Outline of the Talk

•Collection of Twitter Follower Market Data

•Characteristics of Victims and Customers

•Detecting Market Customers


Outline of the Talk





Active Twitter Follower Markets

Market (sorted by order of returned results)

$ for 10K Followers Pyramid?

Newfollow.info $216 YES

Bigfolo.com $91.99 YES

Bigfollow.net $70 YES

Intertwitter.com $65 NO (fake accounts)

Justfollowers.in $95 YES

Twiends.com $169 NO (fake accounts)

Socialwombat.com $49 NO (fake accounts)

Devumi.com $64 NO (fake accounts)

Hitfollow.info $214 YES

Plusfollower.info $214 YES

Buyactivefans.com $40 NO (fake accounts)


We studied the top-five ranked markets

10

• We queried search engines looking for Twitter Follower Markets

• We developed a classifier to determine whether a website is actually selling followers

Different price, depending on the type of followers sold: real followers are more expensive

Market Sizes

We look at tweets advertising the top five markets10% of the all public tweets (3.3 billion tweets), collected over a period of four months


Market Tweets Victims

BigFollow 662,858 90,083

BigFolo 4,732,016 611,825

JustFollowers 302 257

NewFollow 77,865 38,341

InterTwitter 0 0

Total 5,473,041 740,506

11

Detecting Market Victims

We purchased followers from the most popular five markets


Whoever followed us is a victim

In total, we identified 69,222 victims

12

Detecting Market Customers


Get more followers!

Get more followers!

Get more followers!

Get more followers!

13

Detecting Market Customers

We signed up 180 newly-created accounts as market victims


We identified 2,909 market customers

14

Outline of the Talk





Customer CharacteristicsWe compared our set of customers to a set of two million regular users picked at random


Customer Follower Dynamics


Inflation period Deflation period

Customer Follower DynamicsDuring an observation period of one week:

• Spike in Followers ≥ 50 over an hour: 50% Customers, 0.4% Regular

• Steady decrease of followers for ≥ 10 consecutive hours: 60% Customers, 0.05% Regular

• Change of number of followers ≈ 0:0% Customers, 30% Regular


Victim CharacteristicsDifferent strategies for operating markets• Some markets form dense cliques of victims• Some market’s victims follow many customers


Common characteristics of victim accounts:• Victims follow each other• A small fraction of victim accounts (≈20%) gets suspended

19

Outline of the Talk





Follower Dynamics Detection

Three types of features (calculated over a week)•Increase features (1,000 features)

Number of times spike of d followers during an hour

•Decrease features (168 features)Number of times steady decrease of followers for d consecutive hours

•Stationary features (168 features)Number of times followers remained constant for d consecutive hours


We developed a classifier to detect customers in the wild

Follower Dynamics DetectionGround truth: Set of 2,909 customers and 10,000 regular accounts (monitored for a week)


Classifier: Support Vector Machines

10-fold cross validation: 98.4% true positive rate 0.02% false positive rate

Detecting Customers in the Wild

We monitored our set of two million regular accounts for two weeks


We detected 684 customers•Observed only two million accounts•Purchase needs to happen during our observation

23

Analysis of the Identified CustomersThe detected accounts have the expected characteristics of customers•They belong to wanna-be celebrities and small businesses•They do not post interesting content


Twitter fails in detecting customers: 2 out of 684 were suspended

Buying followers does not help in becoming influential (median Klout 45)• A customer with 103,000 followers → same Klout score as me (57)

24

Discussion

Our proposed approach to detect and block market customers could undermine the foundations of Twitter Account Markets


Market operators could adapt, and try to evade detection•Provide followers slowly•Have no control over the unfollow behavior!

25

Conclusions

• We performed a large-scale study of Twitter Follower Markets• We propose techniques to detect market customers• We advocate for Twitter to adopt similar techniques


Questions?


[email protected]

@gianlucaSB

27

Problem: the Dynamic Classifier is DemandingIn the paper, we propose two alternative methods:•A static filter, to discard as many candidates as possible•A static classifier, that uses static profile information to detect customers


System TP rate FP rate

Static Filter 93.7% 63%

Static Classifier 91% 3.3%

Dynamic Classifier 98.4% 0.02%

28

Social Media

Follow the Green: Growth and Dynamics on Twitter Follower Markets