Analyzing the Dynamics of Communication in Online Social ...Communication is the process by which participating individuals create and share information with one another in order to

Analyzing the Dynamics of Communication in Online Social Networks

Munmun De ChoudhurySchool of Computing, Informatics &Decision Systems EngineeringArizona State University

Committee Members

• Dr HariSundaram, Chair

• Dr K. SelçukCandan

• Dr Huan Liu

• Dr Duncan Watts, Yahoo! Research

• Dr DoreeSeligmann, Avaya Labs

2April 15, 2010 Ph.D Proposal Defense

Goal:Understanding the dynamics and impact of our online social interactions

Information

Media

Network

Background and MotivationContributionsCompleted WorkProposed WorkConclusions

April 15, 2010 5

Modern Social Interactional Modes

Facebook

Slashdot

Engadget

Flickr

LiveJournalDigg

YouTubeBlogger

MetaFilterReddit

MySpaceOrkut

Twitter

Ph.D Proposal Defense

Some Social Media Statistics

6

YouTube 139M users; US$200M [Forbes].

Flickr 3.6B images; 50M users.

Facebook 350M+ active users; 1B pieces of content (web links, news stories, blog posts, notes, photos, etc) shared each week.

Twitter 75M users; 600 tweets per second

MySpace 110M monthly active users; 14 B comments on the site.

Digg 3M unique users; $40M.

Engadget 1,887,887 monthly visitors.

Huffington Post 8.9M visitors.

Live Journal 19,128,882 accounts.

April 15, 2010 Ph.D Proposal Defense

7

What are the impacts of such large-scale online social interaction?

Viral Marketing, Advertizing Campaigns

9

What has been the buzz on the release of iPad?

Collaboration, “Wisdom of the Crowds”

11

What have been the political sentiments regarding the 2008 US Presidential election candidates?

Crisis management w.r.t. real-time events

13

Which is the best news source to read about the recent California / Mexico / Arizona earthquake?

14

What are the challenges? • Scale• Growth• Accessibility• Complexity

• Relevance• Utility• Tools

Facebook features ~400 million users300-600 tweets per second!

Twitter developer API’s rate limit: 20,000 requests only!

About a quarter of all tweets are spam (http://mashable.com/)

Time complexity of breadth first search on a social graph: O(|E|+|V|), |V| and |E| in the order of millions…

Are the extracted groups in a community meaningful?

How do we compare across interactions on Facebook and those on LinkedIn?

http://mashable.com/

ContributionsCompleted WorkProposed WorkConclusions

Background and Motivation

Communication is the process by which participating individualscreate and share information with one another in order to reach amutual understanding [Berger and Calabrese, 1975].

Semiotic tradition

Cybernetictradition

Socio-psychological/Socio-cultural

tradition

Addresses interpretation and generation of meaning as a consequence of the communication process [Craig, 1999]

The model of communication is interpreted as a source that produces a message that is passed along a “channel”, to a receiver that interprets the message [von Foerster, 1995]

Considers communication to be the interaction among individuals, that results in the production and re-production of social order [Littlejohn, 2001]

Information or concept

Media or channel

Social engagement or network

Research Questions

1) How do we analyze the diffusion of information or concepts among the communicating individuals?

Time

2) How does the social engagement give rise to emergent collective behavior and evolution of groups and networks?

Time

3) How do we characterize the evolving properties of the media artifacts via which communication takes place?

Thesis Overview

Information Network Media

Overview of Our Work

23

Characterization Models Observational Studies

InformationDiffusion

Topical CommunicationFlow

Diffusion of User Actions

Impact of Homophily onDiffusion

Evolution of Social Communication Networks

Community Dynamics and External Phenomena

Prototypical CommunicationGroups

Multiplex Ties: Evolution and EmpiricalThresholds on Ego-network Size

Rich Media Communication Patterns

Interestingness of Conversations

Connecting Media to Community

Models of User Behavior on Rich Media


Summary of Publications

24

Characterization Models Observational Studies

InformationDiffusion

1. WI 2007 (17%)2. HT 2008 (24%)3. under review at

ACM Tweb

1. SocialCom 2009 (9%)

2. journal submission for ACM Tweb

1. ICWSM 2010 (18%)2. under review at

KDD 20103. journal submission

TBD


1. HT 2008 (24%)2. CIKM 2008 (16%)

1. HT 20092. ACM TOIS

1. Proposed work (target: WWW 2011, ICWSM 2011)

Rich Media Communication Patterns

1. WWW 2009 (11%)2. Journal submission

for ACM TOMCCAP

1. ICME 2009 (22%) 1. Proposed work (target: KDD 2011)


Completed WorkProposed WorkConclusions

Background and MotivationContributions

Rich Media Communication PatternsRelated publications at: WWW 2009, ICME 2009

27

Why do people repeatedly come back to the same You Tube video?


28

They are returning to do more thanwatch the same video


29

We think it is the conversationsaround the video they find interesting


What is a Conversation?

• A conversation associated with an online social media object is a temporally ordered sequence of comments posted by individuals.

Comment

Participant

Timestamp of comment

Media Object

Media Attributes

Conversation

April 15, 2010 30Ph.D Proposal Defense

31

Themes and participants make conversations interesting


An example of an interesting conversation …

33

• Goal:– What causes a conversation to be interesting, that prompts a user to

participate in the discussion on a posted video?

Our Contributions

Approach:

• Detect conversational themes.

• Determine interestingness of participants and interestingness of conversations based on a random walk model.

• Measure the consequence of a conversations.

• Excellent results on a dataset from YouTube.


April 15, 2010 34

Related Work

Analysis of Media Properties

Theme Extraction

in Social Media

Analysis of CommunicationIn Social Media

• Visualization of evolution of tags [Dubinko et al. 2006]• YouTube traffic characterization [Gill et al. 2007]• Analysis of user-generated media [Cha et al. 2007]• Tagging conventions and strategies (YouTube) [Geisler et al. 2007]• Social dynamics in media sharing [Halvey et al. 2007]• Representative views of landmarks [Kennedy et al. 2008]

Cha et al. 2007

Mei et al. 2007

• Predictivity of online chatter [Gruhl et al. 2005]• Analysis of weblog comments [Mishne 2006]• Correlation of communication & stocks [Choudhury et al. 2007]• Predicting sales using blog communication [Liu et al. 2007]• Discussion threads in Slashdot [Kaltenbruner et al. 2007, Gomez 2008]• Video interactions in YouTube [Benevenuto et al. 2008]• Tendency of topic discussions [Zhou et al. 2008]

• Spatio-temporal theme mining [Mei et al. 2005, 2006]• Mining multi-faceted overviews of topics [Ling et al. 2008]• Video topic discovery [Liu et al. 2008]


35

Conversational Themes• Conversational themes are sets of salient topics associated with

conversations at different points in time.

bag of words (chunks,λi,t) over time slices

t1 t2 tQ

……

……Theme models, θj

conversation i


Theme Model

• Temporal Regularization– A word w in the chunk can be attributed either to the textual context λi,t, or

the time slice t

– Smoothness of theme models over time

, ,

, ,1

( ) , .log , | , expi t i t

K

i t j i t TC w j

L C n w p w t d j

time

time

Them

e d

istr

ibu

tio

n

Them

e st

ren

gth

Chunk context Time context Temporal Smoothing


Theme Model (Contd.)

• Co-participation based Regularization

– If several participants comment on a pair of chunks, their theme distributions are likely to be close to each other.

2

2

,, 1

( ) 1 | | .i m

K

i m j i j mc c C j

R C f c f c

( ) (1 ). ( ) . ( )O C L C R C

conversation ci, f(θj|ci)

ωi,m

conversation cm, f(θj|cm)


Interestingness of Conversations

38

ψ 1-ψImpact due toparticipants

Impact due tothemes

Interestingness of participants

Relationship between participants and conversations

Conversational Theme Distribution

Theme Strength

Interestingness of conversations

X

X

X


t─1 t

ψ

1─ψ

April 15, 2010

( ) ( 1) ( 1) ( 1) ( 1) ( 1). . (1 ). . .tq q q q q qdiag C C P T S CI P I C T I


Interestingness of Participants

39

β1-β

Past communication Preference

April 15, 2010

1( ) ( 1) ( 1)

-1 ( 1) ( 1) ( 1) ( 1) ( 1) ( 1)1 2 3

(1 ). . . ,

where . . . . . .

qq q q

q q q q q q q

p T S

L P F P C C

I A P T

A P I P I P I


40

Interestingness of participants and conversations mutually reinforce each other


41

Joint Optimization of Interestingness

• A joint optimization framework, which maximizes the two interestingness measures for optimal X=(α1, α2, α3,Ψ) and also incorporates temporal smoothness:

2 2

1( ) . (1 ). exp expP Cg d dP CX I X I X

Interestingness of participants


Regularization of participants’ interestingness

Regularization of conversations’ interestingness


42

Three consequence metrics of interestingness

t t+δ

acti

vity

Participant activity

t t+δ

participant cohesiveness

Participant Cohesiveness

t t+δ

conversation

Thematic Interestingness


43

YouTube Dataset

• ‘News & Politics’ category on YouTube – rich communication on highly dynamic events.– 132,348 videos

– ~ 9M unique participants

– ~ 89M comments

– 15 weeks from June 20, 2008 to September 26, 2008


45

Analysis of Interestingness of Participants

Interestingness of participants is less affected by number of comments during significant external events


46

Analysis of Interestingness of Conversations

Mean interestingness of conversations increases during periods of several external events; however, certain highly interesting conversations always occur at different weeks irrespective of events.


47

Evaluation using Consequences

• Interestingness is computed using five techniques –– our method with temporal smoothing (I1), – our method without temporal smoothing (I2) and – the three baseline methods,

• B1 (comment frequency), • B2 (novelty of participation), • B3 (co-participation based PageRank).


April 15, 2010 48

Summary

• Summary– Why do people repeatedly come back to

the same YouTube video?

– Developed a probabilistic framework that characterizes “interestingness” of conversations

– Our method can explain future consequences

• Future Work– User subjectivity

– Personalized recommendations

t t+δ

participant cohesiveness

Participant Cohesiveness


Information DiffusionRelated publications at: WI 2007, HT 2008, SocialCom 2009, ICWSM 2010, KDD 2010 (under review), ACM Tweb (under review)

XKCD: Seismic Waves

XKCD: Seismic Waves #iPad

People from the south-west of US?

Apple gadget lovers?

What is homophily?

• The tendency of individuals to bond more with ones who are “similar” to them, compared to ones who are “dissimilar” *McPherson et al. 2001+.

• Observed sociological consequence: similarity breeds connection.

52

Information roles

LocationActivity behavior


Twitter

• “Tweets”: 140 character length shared content.– RT (or re-tweet feature), hashtags (e.g. #iranelection), bit.ly encoded URLs

• Follower / Following relationship.

• “Trending topics” e.g. #musicmonday, #formulaone.

April 15, 2010 53

• Diffusion via (1) RT feature, (2) shared URL (e.g. bit.ly, tinyurl), (3) same hashtag

(1) RT based diffusion

(2) URL based diffusion

(3) hashtag based diffusion


Diffusion characteristics vary across different user attributes along which the diffusion process takes place

Our Contributions

• Goal:

– How does attribute homophily among users affect the diffusion of information on online social media?

55

• Approach:– Representation of diffusion process via a

structure called “diffusion series”

– A DBN based probabilistic model that can predict diffusion characteristics on an attribute graph

• Excellent results on a large dataset from Twitter indicating that homophily affects diffusion, but depends on topic and metric of measurement.


Related Work

April 15, 2010 56

Studies on Homophily

Diffusion Dynamics

Homophily &Social contagion

Inter-connectedness between homogeneous composition of groups and emergent homophily [McPherson et al. 2001]

Topic diffusion in the Blogosphere [Gruhl et al. 2004]

Diffusion of “gestures” on Second Life *Bakshy et al. 2009]

Contagion via Facebook “News Feed” *Sun et al. 2009]

Homophily and contagion over romantic and sexual networks [Bearman et al. 2004]

Compliance and conformity in networks [Cialdini et al. 2004]


April 15, 2010 Ph.D Proposal Defense 57

Data Model

Diffusion Series

58

Social graph Attribute social graph

Diffusion series


Characterizing Diffusion

• User-based(Volume, participation, dissemination)

• Topology-based(Reach, spread, cascade instances, collection size)

• Time-based(Rate)

59

Volume

ParticipationDissemination

Reach

Spread

Cascade Instances

Collection Size

Rate



Attribute social graph at tN

Social actions of users until tN

Predict the users likely to perform the social action

at tN+1

Construct diffusion

series at tN+1

Predict diffusion characteristics at tN+1

Does a particular chosen attribute predict the diffusion characteristics well? That is, does homophily on the particular attribute impact the diffusion process?

Method

tN tN+1

Prediction Framework

61

• Dynamic Bayesian network based representation of user’s social actions (i.e. “tweeting” behavior in this work)

– User context (e.g. past history of activity, degree of activity of ego-network, popularity of topic) i.e. Fi,N, states i.e. Si,Nand observed action i.e. Oi,N

• Two latent states: “vulnerable or socially aware” (Si=1) and “indifferent or socially unaware” (Si=0)

• Goal: estimate the probability of observed social action at a future time tN+1, based on the latent states Si,Nand the contextual attributesFi,N


Estimation

62

, 1

, 1

, 1 , , , 1 , 1 , , , 1 , ,

, 1 , 1 , 1 , ,

| , | , , | ,

| | , ,

i N

i N

i N i N i N i N i N i N i N i N i N i NS

i N i N i N i N i NS

P O O P O S O P S O

P O S P S S

F F F

F

Estimate user context

Estimate probability of user state given context and past state

Estimate probability of user action given the stateA Hidden Markov Model

where the actions are the emissions; use Viterbi to predict the likely sequence

Multinomial density of states over the contextual attributes with a Dirichlet prior

where,Oi,N= action of user iat time slice tN

Fi,N= context of user iat time slice tN

Si,N= state of user iat time slice tN

April 15, 2010

Predicting User States

63

…..

…..

…..

, 1 , , , , , 1 ,

Using Bayes rule and first order Markov property,

( | , ) ( | ). ( | )i N i N i N i N i N i N i NP S S P S P S S F F

multinomial

Dirichlet prior

, ;

, 1 , , , , , 1 ,

, , , 1 , 1

, ;

, ; ,

, ; , 1

Parameter estimation using MAP:

( | , ) log( ( | )) log( ( | ))

log ( ; ) log ( ; )

!1

log log! ( )

i N jk

i N i N i N i N i N i N i N

i N i N i N i N

i N jkjk

i N jk i Njki N jk i N

jk

L S S P S P S S

S

SB

F

F F

multinom F Dirichlet

F

F

, ;

1

, 1 , 1

,

where ( ) is a beta-function with the parameter .

i N jlS

jl

i N i NB



Attribute social graph at tN

Social actions of users until tN

Predict the users likely to perform the social action

at tN+1

Construct diffusion

series at tN+1

Predict diffusion characteristics at tN+1

How we do we evaluate?

tN tN+1

1 1 1

1 11

1 1

1

1 ( ( ), ( )) where, ( ) is the search volume, and

ˆ( ) | ( ( ))|/ ,

ˆ ˆ| ( ( ))| is the number of nodes at slot in the collection ( ), and

ˆQ | ( ( ))|. Si

D S SN N N

DN m N D

m N

m N m N

D m Nm

D E E E

E l S Q

l S l S

l S

milarly,

Distortion Measurement• Saturation measurement, i.e. ability of a given attribute social graph to predict

user-based, topological and temporal diffusion characteristics at a future time slice, given using the Kolmogorov-Smirnov (KS) statistic:

65

• Utility measurement, i.e. ability of a given attribute social graph to correlate with external temporal variables like user search behavior and news items featured online (http://news.google.com/), given as:

1 1 11 ( ( ), ( )) where, ( ) is the news volume.DN N ND E E E


1 1 1

1

ˆ ˆ1 ( ( ), ( )) where, ( ) is the predicted value of the

-th diffusion characteristic at .

M M MN N N

N

D E E E

M t

http://news.google.com/


Experimental Studies

Experimental Setup

• ~465K users, ~836K edges (“follower” / “following” relationships) and 29.5M tweets.

• 125 randomly chosen “trending topics” from Twitter, between Oct and Nov 2009.

• Trending topic – theme association based on OpenCalais (http://www.opencalais.com/).

April 15, 2010 67

Themes Trending topics

Politics Obama, Senate, Afghanistan, Tehran, Healthcare

Entertainment_Culture Beyonce, Eagles, Michael Jackson, #britney3premiere

Sports Chargers, Cliff Lee, Dodgers, Formula One, New York Yankees

Technology_Internet Android 2, Bing, Google Wave, Windows 7, #Firefox5

Social Issues Swine Flu, Unemployment, #BeatCancer, #Stoptheviolence


• Dataset released for non-commercial research purposes:

http://www.public.asu.edu/~mdechoud/temp/released-data/

http://www.opencalais.com/




Temporal Analysis

68

Best: IRO Best: CCR

Best: ACTBest: ACT

Topical Analysis

69

A: Business-Finance, B: Politics, C: Entertainment-Culture, D: Sports, E: Technology-Internet, F: Human Interest, G: Social Issues, H: Hospitality-Recreation.

• For a given theme, the attribute that yields best prediction depends on:

Effect of external events

Habitual properties of users


Baseline Models

• GenModel: Generative model

• Cascade: Information cascades

• LinRegress: Linear regression

• DegAct: Degree of activity

• Random: Random sample

70

GenModel

Cascade

, 1 1 , ;1 2 , ;2 , ;( ) . . .i N i N i N k i N kP O F F F


Comparative Study

71

Summary

• Contributions:– Quantified impact of homophily on the diffusion

process in online social media.

– Proposed and evaluated a computational model that predicts diffusion characteristics over time given an attribute social graph.

– Excellent results on a large dataset from Twitter.

• Future directions:– What is the relationship between diffusion process

and social contagion?

– Can we develop computational models that maximize predicted diffusion characteristics given several homophily attributes?


Proposed WorkConclusions

Background and MotivationContributionsCompleted Work


Related publications at: HT 2008, CIKM 2008, HT 2009, ACM TOIS

Online communication networks are multi-modal and dynamic

tim

e

“active” network only limited to a few individuals and changes over time

Network evolves based on microscopic interactions at the ego-centric level

Proposed Goal

• Evolution of multiplex networks:– What are appropriate representations of multiplex

networks based on multi-relational interactions on online social media?

• Ego-network size thresholds:– Are there empirical thresholds on the sizes of ego-

networks corresponding to a given modality?

– Do dependencies exist among the thresholds across these different modalities?

– Can we model these inter-dependencies among the thresholds of multiple modalities to develop a generalized model of evolution of ego-networks?

April 15, 2010 76

tim

ePh.D Proposal Defense

Who can benefit from this research?

A news reporter A political analyst A company The average Joe…

Prior Work

April 15, 2010 78

MultiplexNetworks

Social NetworkEvolution

Thresholds ofEgo-networks

1. Multiplexity as a fundamental aspect of social relations [Dacin et al., 1999, Robins et al., 2005]2. Consideration to the different types of resources transferred through linkages [Lomi and Pattison, 2006, Robins et al., 2007]3. Structural embeddedness[Granovetter 1973, Uzzi, 1997, Gulati, 1999, Gulati and Gargiulo, 1999]

1. Cognitive limits on the number of friends an individual [Dunbar, 1992,Dunbar, 1993]2. Facebook study on applicability of Dunbar’s number (http://www.insidefacebook.com/2009/02/27/facebooks-in-house-sociologist-shares-stats-on-users-social-behavior/)

1. Preferential attachment [Barabasi and Albert, 1999, Abello et al., 2002,Cooper and Frieze, 2003]2. Copying model [Kleinberg et al., 1999, Kumar et al., 2000]3. Diameter and network densification [Leskovec et al., 2005, Leskovecet al., 2008]


http://www.insidefacebook.com/2009/02/27/facebooks-in-house-






Research Steps


1 2

1 2

( )

( ,{ , , , }; ), directed social graph at a certain time , where there are

sets of multiplex ties { , , , } between users in .

implies that communicates with on modality at

n

n

kij k i j

G V E E E t t

n E E E V

e E u u k

time .t

Stacked representationof a multiplex network at time t given as:

• Research challenges:– predict the ego-network of each user uiin V at t+1

for modality k, given data until time t.

– derive the network G’(V,{E1, E2, …, En};t+1) for all nmodalities at t+1.

– determine threshold d(k)t+1 corresponding to all

users uiin Vthat bounds the size of the ego-networks of users for modality k at t+1.

– determine the conditional relationship between d(k)

t+1 and d(l)t+1 where k, l ≤n

1 2( ,{ , , , }; )nG V E E E t

Discussion

• Evaluation strategies– Compare with observed multiplex social ties

at a future time slice

– Ethnographic studies of evaluation of thresholds of ego-network sizes

• Potential datasets for evaluation– Digg dataset: features multiple

interactional modalities – “digging”, commenting and replying

– Flickr dataset: diverse rich media interactions – photo upload, commenting, tagging, “favoriting”, group /photo pool subscription

– Others: YouTube, Twitter (?)


Photo upload

tagging

Photo pool association

Commenting

Tentative Timeline

81

1. Proposed Work I: Evolution of ego-networks over multiplex ties:2010/08 – 2011/02• Background and lit survey: 2010/08• Problem formulation: 2010/08 – 2010/09• Data collection: 2010/09• Experimental studies: 2010/10• Conference submission: 2010/11• Journal submission: 2010/12 – 2011/02

2. Proposed Work II: Models of user behavior on rich media:2010/11 – 2011/04• Background and lit survey: 2010/11• Problem formulation: 2010/11 – 2010/12• Data collection: 2010/12• Experimental studies: 2011/01• Conference submission: 2011/02• Journal submission: 2011/03 – 2011/04

3. Thesis writing:2011/02 – 2011/054. Dissertation defense:2011/05


Conclusions

Background and MotivationContributionsCompleted WorkProposed Work

Information Network Media

Conclusions

• Goal:– Analyzing the dynamics and impact of online

social interactions

• Completed Work:– Characterization of rich media artifacts, via

interestingness property• [YouTube dataset] Interesting conversations on rich social

media sites show consequential impact on sociological behavior

– Information diffusion and its relationship to attribute homophily• [Twitter dataset] Diffusion characteristics vary across

topics and is affected by homophily

• Proposed Work:– Evolution of ego-networks over multiplex social

ties


Open Issues & Future Directions

• Social data sampling i.e. representativeness of observations and models

• Modeling sociological behavior at smaller topological / temporal scales i.e. in the “long tail”

• Application of proposed methods to complex real-time communication contexts, e.g. in enterprise business processes, crisis mitigation teams, financial domains

• Ethnographic and longitudinal studies to incorporate user feedback on the design and analysis of social communication systems


Thanks!

[email protected] Publications / Datasets: http://www.public.asu.edu/~mdechoud/

Questions?

mailto:[email protected]

http://www.public.asu.edu/~mdechoud/

Analyzing the Dynamics of Communication in Online Social ...Communication is the process by which participating individuals create and share information with one another in order to

Documents