Citizen Sensing, Social Media Analytics, and Applications

Citizen Sensor Data Mining,
Social Media Analytics and
Development Centric Web Applications.
Tutorial at
Semantic Technology Conference,
San Francisco, CA.

Karthik Gomadam Amit Sheth Selvam Velmurugan
Accenture Technology Labs, Kno.e.sis @ eMoksha, Kiirti
San Jose Wright State
University

Monday, June 6, 2011

Meena Nagarajan Selvam Velmurugan
(Content Analysis) (Kiirti, eMoksha NGOs)

Hemant Purohit
Amit Sheth
(People & Network analysis)
(Semantic Web)
Ashutosh Jadhav
(Event Analysis)

Lu Chen
Pramod Anantharam (Sentiment Analysis)
(Social & Sensor web) Pavan Kapanipathi
(Real Time Web)


A Quick Word

Much of the work discussed in this tutorial is
primarily the doctoral research by Dr. Meena
Nagarajan, currently at IBM Almaden. It also
includes current work done at kno.e.sis center at
Wright State University.


Outline

Citizen Sensing: Role, Enablers, Apps
Systematic Study Social Media
Citizen Sensing @ Real-‐‑time
Emerging Research Areas
‣ Spam and Trust in Social Media, Mobile Social Computing
Research Application: Twitris
Tutorial part 2


Citizen Sensing

Everyday users of Web2.0 and social networks:
Citizens of an Internet- or Web-enabled social
community
Observation and Information reported by citizens
=> Citizen Sensing
Human-in-the-loop (participatory) sensing + Web
2.0 + mobile computing = emergence of
" citizen-sensor networks


Social Signals

The activity of observing, reporting, disseminating
information via text, audio, video and built in device
sensor (and smart devices),
‣ Creating social signals through aggregation, enhancement,
analysis, visualization, and interpretation.

Immense potential to disseminate information
quickly and in real-time


Enablers: Mobile Devices &
Ubiquitous Connectivity
Mobile device fast emerging as our primary tool
‣ Redeﬁnes the way we engage with people, information,
etc.
Global, Ubiquitous, always available
Sense where you are, how you are, …



Global, Ubiquitous, always available


Mobile Platforms Hit Critical Mass
‣ Over 5 billion users
‣ 1+B with internet connected mobile devices (2010)
‣ Smartphones > Notebooks + Netbooks (2010E)
‣ 500K+ mobile phone applications
‣ 74% of mobile phone users (2.4B) worldwide texted (2007)


Enablers: Web 2.0 & Social Media

500M+ Facebook Users
100M+ Twitter users, 85M+ tweets/day
Internet Users: 1.8 Bln
Content dissemination medium
‣ Even for traditional media (@cnn, @nytimes)



100M+ Twitter users, 85M+ tweets/day



Types of UGC: Twitter(text/microblogs), Facebook
(multimedia),YouTube(videos),
Flicker(images), Blogs(text),
Ping: (Social network for music)



Flicker(images), Blogs(text),


Citizen Sensors in Action

Iran election
Haiti Earthquake
US healthcare
debate


Revolution 2.0
Political/Social Activism
“If you want to liberate a government, give them the
internet.” - Wael Ghonim (Egyptian social activist)
When Blitzer asked “Tunisia, then Egypt, what’s
next?,” Ghonim replied succinctly “Ask Facebook.”


Revolution 2.0

When Blitzer asked “Tunisia, then Egypt, what’s
next?,” Ghonim replied succinctly “Ask Facebook.”


Revolution 2.0


Citizen Journalism

Twitter Journalism


Social Media Inﬂuence:
Intelligence, News & Analysis
Many media companies use Facebook and Twitter
as news-delivery platform. Many individuals rely on
them as news source. News is increasingly social.


Business Intelligence Trend
SpoTing, Forecasting, Brand
Tracking and Crisis Management
Sysomos : http://www.sysomos.com/
Trendspotting : http://trendspotting.com
Simplify : http://simplify360.com/
Shoutlet : http://www.shoutlet.com/
Reputation (Defender) :
http://www.reputationdefender.com/


Development
(Education, Health, eGov)
LiveMocha (http://www.livemocha.com/)
‣ Online Language learning tool with social engagement
‣ bridging the gap!!
Soliya (http://www.soliya.net/)
‣ Dialogue between students from diverse " backgrounds
across the globe using latest multimedia technologies
Project Einstein (http://digital-democracy.org/what-we-do/programs/)
‣ A photography-based digital penpal program connecting
youths in refugee camps to the world


Development

Soliya (http://www.soliya.net/)
‣ Dialogue between students from diverse " backgrounds
across the globe using latest multimedia technologies


Development



Development
PatientsLikeMe (http://mashable.com/2010/07/13/social-media-health-trends/)
TrialX (http://trialx.com)

Image: hMp://www.dragonsearchmarketing.com/
blog/
social-‐‑media-‐‑development-‐‑through-‐‑visual-‐‑aids-‐‑
tools/


Why People-‐‑Content-‐‑Network
metadata?


Dimensions of Systematic Study
of Social Media

Spatio - Temporal -Thematic
+
People - Content - Network


Social Information
Processing
"Who says what, to whom, why, to what extent
and with what effect?" [Laswell]
Network: Social structure emerges from the
aggregate of relationships (ties)
People: poster identities, the active effort of
accomplishing interaction
Content : studying the content of ommunication.


Studying Online Human Social
Dynamics
How does the (semantics or style of) content ﬁt
into the observations made about the network?
‣ Often, the three-‐‑dimensional dynamic of people,
content and link structure is what shapes the social
dynamic.


Dynamics


Dynamics
Example: how does the topic of discussion,
emotional charge of a conversation, the presence of an
expert and connections between participants; together
explain information propagation in a social network?


Metadata/Annotations

Metadata: an organized way to study
‣ types

‣ creation/extraction and storage

‣ use


The Anatomy of a Tweet


People Metadata: Variety of
Self-‐‑expression Modes on Multiple
Social Media Platforms
Explicit information from user proﬁles
‣ User Names, Pictures, Videos, Links, Demographic
Information, Group memberships...
‣ Often is not updated
Implicit information from user a+ention metadata
‣ Page views, Facebook 'ʹLikes'ʹ, Comments; TwiMer
'ʹFollows'ʹ, Retweets, Replies..


People Metadata: Various Levels

Demographic

Interests

Activity

Network


People Metadata: Continued
User Demographic Metadata Interest Level Metadata
•User-id •Author type
•Screen/Display-name of •Trustee/donor, journalist,
user blogger, scientist etc.
•Real name of user • Favorite tweets
•Location • Types of lists subscribed
•Proﬁle Creation Date • Style of Writing –
•User description personality indicator
•User Bio • No. of Followees
•URL • Author type trend of
Followees


People Metadata: Continued
Activity Level Metadata Influence Level Metadata
(Inferring People Metadata from Network level Information)

•Age of the profile •No. of Followers – normal, influential
•Frequency of posts •No. of Mentions
•Timestamp of last status •No. of Retweets/Forwards
•No. of Posts •No. of Replies
•No. of Lists/groups created •No. of Lists/groups following
•No. of Lists/groups subscribed •No. of people following back

•Authority & Hub Scores
Web Presence:
•User affiliations
•KLOUT Score – influence measure (www.klout.com)

Content Metadata

Content Independent metadata
‣" date, location, author etc
Content Dependent metadata
‣ Direct content-based metadata
‣ Explicit/Mentioned Content metadata
‣ named entities in content

‣ Implicit/Inferred Content Metadata
‣ related named entities from knowledge sources

‣ Indirect content-based metadata (External metadata)
‣ context inferred from URLs in content (images, links to articles,
FourSquare checkins etc.)

Content Metadata

Content Dependent metadata
‣ Direct content-based metadata
‣ Explicit/Mentioned Content metadata
‣ named entities in content

‣ Implicit/Inferred Content Metadata
‣ related named entities from knowledge sources

‣ Indirect content-based metadata (External metadata)
‣ context inferred from URLs in content (images, links to articles,
FourSquare checkins etc.)

Content Metadata


Content Independent Metadata

For Tweets
‣ Published date and time
‣ Location (where tweet was generated from)
‣ Tweet posting method (smart-phone, twitter.com,
clients for twitter)
‣ Author information



For Text messages
‣ Published date and time
‣ Origin location
‣ Recipient
‣ Carrier information


Content Dependent Metadata (Tweet)
Direct Content-‐‑based Metadata
Direct Content-based Metadata

Indirect content-based metadata (External metadata)


Content Dependent Metadata
Direct Content-‐‑based Metadata


Network Metadata

Connections/Relationships (foundation for the network)
matter!
Structure Level Metadata Relationship Level Metadata

•Community Size •Type of Relationship
•Community growth rate •Relationship strength
•Largest Strongly Connected •User Homophily based on
Component size certain characteristic (e.g.,
•Weakly Connected Components Location, interest etc.)
& Max. size •Reciprocity: mutual relationship
•Average Degree of Separation •Active Community/ Ties
•Clustering Coeﬃcient

Metadata: Creation, Extraction
and Storage


Metadata Creation & Extraction

Extracted Metadata
‣ Directly visible information from the user proﬁle, tweet
content & community structure
Created Metadata
‣ After processing information in the user proﬁle, content
and/or network structure


An Example

Length: 144 characters; General topic: Egypt protest
This poor {sentiment_expression: {target:”Lara
Logan”, polarity:”negative”}} woman! RT @THR CBS
News'{entity:{type=”News Agency”}} Lara Logan
{entity:{type=”Person”}} Released From Hospital
{entity:{type=”Location”}} After Egypt{entity:
{type=”Country”} Assault{type=”topic”}
http://bit.ly/dKWTY0 {external_URL}


Why Semantic Web is a standard
for social metadata?

Rich Snippet, RDFa, open graph, semantic web
based social data standards
Relationships/connections play central role
‣ Relationships as ﬁrst class object is important


Semantic Web: A Very Short
Primer


Primer
Representation
‣ RDF
‣ relationships as ﬁrst class object <subject,
predicate,object>
‣ OWL
‣ Representing Knowledge and Agreements:
nomenclature, taxonomy, folksonomy, ontology


Primer
Annotation
‣ RDFa, Xlink, model reference


Primer
Annotation
Web of Data
‣ Linked Open Data


Primer
Annotation
Web of Data
‣ Linked Open Data
Querying
‣ SPARQL; Rules: SWRL, RIF


How to save and use metadata?

Store metadata as data and use standard database
techniques
Use ﬁltering and clustering, summarization,
statistics - implicit semantics



Use ﬁltering and clustering, summarization,
statistics - implicit semantics



Use explicit semantics and Semantic Web
standards and technologies
‣semantics = meaning
‣richer representation, support for relationships, context
‣supports use of background knowledge
‣better integration, powerful analysis
Semantics- the implicit, the formal and the
powerful
Social metadata on the Web


Metadata Extraction from
Informal Text
Meena Nagarajan, Understanding User-Generated Content on
Social Media, Ph.D. Dissertation, Wright State University, 2010


Characteristics of Text on Social
Media


The Formality of Text


Content Analysis-‐‑Typical Sub-‐‑tasks

Recognize key entities mentioned in content
‣ Information Extraction (entity recognition, anaphora
resolution, entity classiﬁcation..)
‣ Discovery of Semantic Associations between entities
Topic Classiﬁcation, Aboutness of content
‣ What is the content about?
Intention Analysis
‣ Why did they share this content?



Topic Classiﬁcation, Aboutness of content
‣ What is the content about?
Intention Analysis



Intention Analysis



Sentiment Analysis
‣What opinions are people conveying via the content?
Author Proﬁling
‣What can we infer about the author from the content he
posts?
Context (external to content) extraction
‣URL extraction, analyzing external content


Research Eﬀorts, Contributions in
this space..
Examining usefulness of multiple context cues
for text mining algorithms
‣ Compensating for for informal, highly variable
language, lack of context
‣ Using context cues: Document corpus, syntactic,
structural cues, social medium, external domain
knowledge…
In this talk, highlighting sample metadata
creation tasks: NER, Key Phrase Extraction,
Intention, Sentiment/Opinion Mining


Part 1. NER, Key
Phrase Extraction
Named Entity Recognition
‣ I loved <movie> the hangover </movie>!
Key Phrase Extraction


Multiple Context Cues Utilized for
NER in Blogs and MySpace


Multiple Context Cues Utilized for
Keyphrase Extraction from TwiTer,
Facebook and MySpace


Focus, Impact

Techniques focus on
‣ relatively less explored content aspects on social
media platforms
Combination of top-down, bottom-up analysis
for informal text
‣ Statistical NLP, ML algorithms over large corpora
‣ Models and rich knowledge bases in a domain


NAMED ENTITY
RECOGNITION


NAMED ENTITY
RECOGNITION
I loved your music Yesterday!
“It was THE HANGOVER of the year..lasted
forever..
So I went to the movies..badchoice picking “GI
Jane”worse now”


NAMED ENTITY
RECOGNITION

Identifying and classifying tokens


NER in prior work vs. NER for
Informal Text


Cultural Named Entities

NER focus in this work: Cultural Named
Entities
Artifacts of Culture
‣ Name of a books, music albums, ﬁlms, video games,
etc.
Common words in a language
‣ The Lord of the Rings, Lips, Crash, Up, Wanted,
Today, Twilight, Dark Knight…


Characteristics of Cultural Entities

Varied senses, several poorly documented
‣ Merry Christmas covered by 60+ artists Star Trek:
movies, TV series, media franchise.. and cuisines !!
Changing contexts with recent events
‣ The Dark Knight reference to Obama, health care
reform
Unrealistic expectations
‣ Comprehensive sense deﬁnitions, enumeration of
contexts, labeled corpora for all senses ..
‣ NER Relaxing the closed-world sense assumptions


NER in prior work vs.
NER for Informal Text


A Spot and Disambiguate
Paradigm
NER generally a sequential prediction problem
‣ NER system that achieves 90.8 F1 score on the
CoNLL-2003 NER shared task (PER, LOC, ORGN
entities) [Lev Ratinov, Dan Roth]
Focus of approach: Spot and Disambiguate
Paradigm
Starting off with a dictionary or list of entities we
want to spot


A Spot and Disambiguate
Paradigm
Spot, then disambiguate in context (natural
language, domain knowledge cues)
Binary Classiﬁcation
Is this mention of “the hangover” in a sentence
referring to a movie?


Algorithmic Contributions
Supervised Algorithms


Algorithmic Contributions
Supervised Algorithms
Examples:
“I am watching Pattinson scenes in <movie
id=2341> Twilight</movie> for the nth time.”
“I spent a romantic evening watching the Twilight
by the bay..”
“I love <artist id=357688>Lily’s</artist> song


Multiple Senses in the Same
Domain


Algorithm Preliminaries

Problem Defn
‣ Cultural Entity Identiﬁcation : Music album, tracks
‣ Smile (Lilly Allen), Celebration (Madonna)
Corpus: MySpace comments
‣ Context-poor utterances
" “Happy 25th Lilly, Alﬁeis funny”



Corpus: MySpace comments
‣ Context-poor utterances



Goal: Semantic Annotation of
music named entities (w.r.t
MusicBrainz)

Using a Knowledge Resource for
NER is not straight-‐‑forward..


Approach Overview

Scoped Relationship graphs
‣Using context cues from the
content, webpage title, url…
new Merry Christmas tune
‣Reduce potential entity spot size
new albums/songs
‣Generate candidate entities
‣Spot and Disambiguate


Sample Real-‐‑world Constraints

Career Restrictions
‣“release your third album already..”
Recent Album restrictions
‣“I loved your new album..”
Artist age restrictions
‣”happy 25th rihanna, loved alﬁe btw..” etc.


Non-‐‑Music Mentions

Challenge 1: Several senses in the same domain
‣ Scoping relationship graphs narrows possible senses
‣ Solves the named entity identiﬁcation problem
partially

Challenge 2: Non-music mentions
‣ Got your new album Smile. Loved it!
‣ Keep your SMILE on!
" " " "
" " " "


Using Language Features to
eliminate incorrect mentions..
Syntactic features
‣ POS Tags, Typed dependencies..
‣ Example here
Word-level features
‣ Capitalization, Quotes
Domain-level features


Supervised Learners


Hand Labeling -‐‑ Fairly Subjective

1800+ spots in MySpace user comments from
artist pages
Keep your SMILE on!

–good spot, bad spot, inconclusive?

4-‐‑way annotator agreements

–Madonna 90% agreement

–Rihanna 84% agreement

–Lily Allen 53% agreement


Dictionary SpoTer + NLP Step

Daniel Gruhl, Meena Nagarajan, Jan Pieper, Christine Robson, Amit Sheth, Context and Domain
Knowledge Enhanced Entity SpoMing in Informal Text, The 8th International Semantic Web Conference,
2009: 260-‐‑276


NER on Social Media Text using
Domain Knowledge
Highlights issues with using a domain
knowledge for an IE task
Two stage approach: chaining NL learners over
results of domain model based spotters
Improves accuracy up to a further 50%
‣ allows the more time-intensive NLP analytics to
run on less than the full set of input data


BBC SoundIndex (IBM Almaden):
Pulse of the Online Music

" "

Daniel Gruhl, Meenakshi Nagarajan, Jan Pieper, Christine Robson, Amit Sheth: “Multimodal Social
Intelligence in a Real-‐‑Time Dashboard System,” special issue of the VLDB Journal on "ʺData Management
and Mining for Social Networks and Social Media"ʺ, 2010 CHECK hMp://www.almaden.ibm.com/cs/
projects/iis/sound/


The Vision

http://www.almaden.ibm.com/cs/projects/iis/sound/


Several Insights
Trending popularity of artists Trending topics in artist pages

Only 4% -‐‑ve sentiments, perhaps ignore the Sentiment Ignoring Spam can change ordering
Annotator on this data source? of popular artists


Predictive Power of Data

Billboards Top 50 Singles chart during the week of
Sept 22-28 ’07 vs. MySpace popularity charts.
User study indicated 2:1 and upto 7:1 (younger age
groups) preference for MySpace list.
Challenging traditional polling methods!


Key Phrase Extraction: Example

Key phrases extracted from prominent discussions
on Twitter around the 2009 Health Care Reform
debate and 2008 Mumbai Terror Attack on one day


Key Phrase Extraction from SM
Text
Different from Information Extraction
Extracting vs. Assigning Key Phrases " Focus:
Prior work focus: extracting phrases that
summarize a document -- a news article, a web
page, a journal article, a book..
Focus: summarize multiple documents (UGC)
around same event/topic of interest


Key Phrase Extraction on SM
Content
Focus: Summarizing Social Perceptions via key
phrase extraction
Preserving/Isolating the social behind the social
data
‣"What is said in Egypt vs. the USA should be viewed in
isolation


Key Phrase Extraction on SM
Content
‣ Accounting for redundancy, variability, off-topic
content
" “Met up with mom for lunch, she looks lovely as ever,
good genes .. Thanks Nike, I love my new
Gladiators ..smooth as a feather. I burnt all the calories of
Italian joy in one run.. if you are looking for good Italian
food on Main, Bucais the place to go.”


Social and Cultural Logic in SMC

Thematic components
‣ similar messages convey similar ideas
Space, time metadata
‣ role of community and geography in communication
Poster attributes
‣ age, gender, socio-economic status reﬂect similar
perceptions


Feature Space (common to several
eﬀorts)
Focus: n-grams, spatio-temporal metadata (social
components)
Syntactic Cues: In quotes, italics, bold; in
document headers; phrases collocated with
acronyms


Feature Space (common to several
eﬀorts)
Document and Structural Cues: Two word
phrases, appearing in the beginning of a
document, frequency, presence in multiple similar
documents etc.
Linguistic Cues: Stemmed form of a phrase,
phrases that are simple and compound nouns in
sentences etc.


Key Phrase Extraction: Overview

“President Obama in trying to regain control of the
health-care debate will likely shift his pitch in
September”
" 1-grams: President, Obama, in, trying, to, regain, ...
" 2-grams: “President Obama”, “Obama in”, “in
trying”, “trying


A descriptor is an n-gram weighted by:

‣ Thematic Importance
‣ TFIDF, stop words, noun phrases
‣ Redundancy: statistically discriminatory in nature
‣ variability: contextually important
‣ Spatial Importance (local vs. global popularity)
‣ Temporal Importance (always popular vs. currently trending)


Eliminating Off-topic Content [WISE2009]
Frequency based heuristics will not eliminate
off-topic content that is ALSO POPULAR


Approach Overview

“Yeah i know this a bit off topic but the other
electronics forum is dead right now. im looking
for a good camcorder, somethin not to large that
can record in full HD only ones so far that ive
seen are sonys”
“CanonHV20.Great little cameras under $1000.”


Approach Overview

Assume one or more seed words (from domain
knowledge base) C1 -['camcorder']
Extracted Key words / phrases
C2 -['electronics forum', 'hd', 'camcorder', 'somethin', 'ive',
'canon', 'little camera', 'canon hv20', 'cameras', 'offtopic']

Gradually expand C1 by adding phrases from C2
that are strongly associated with C1
Mutual Information based algorithm [WISE2009]


Key Phrases and Aboutness
Evaluations
Are the key phrases we extracted topical and
good indicators of what the content is about?
‣ If it is, it should act as an effective index/search
phrase and return relevant content
Evaluation Application: Targeted Content
Delivery


Targeted Content
Delivery -‐‑Evaluations
12K posts from MySpace and Facebook
Electronics forums
‣ Baseline phrases: Yahoo Term Extractor
‣ Our method phrases: Key phrase extraction,
elimination
Targeted Content from Google AdSense


Targeted Content for all content
vs. extracted key phrases


User Studies and Results


Impact and Contributions

TFIDF + social contextual cues yield more useful
phrases that preserve social perceptions
Corpus + seeds from a domain knowledge base
eliminate off-topic phrases effectively


Intention Mining


Targeted Content Delivery via

Intention Mining
On social networks
Use case for this talk
‣" Targeted content = content-based " advertisements
‣ " Target = user proﬁles
Content-based advertisements CBAs
‣" Well-known monetization model for online content


Circa. 2009 Content-‐‑based Ads


Circa. 2009 -‐‑Ads on Proﬁles


What is going on here

Interests do not translate to purchase intents
‣" Interests are often outdated..
‣ " Intents are rarely stated on a proﬁle..
Cases that do seem to work
‣" New store openings, sales
‣ " Highly demographic-targeted ads


Intents in User


Content Ads Outside Proﬁles


Targeted Content-‐‑based
Advertising
Non-trivial

‣ Non-policed content
Brand image, Unfavorable sentiments

‣ People are there to network
User attention to ads is not guaranteed

‣ Informal, casual nature of content
‣ People are sharing experiences and events
Main message overloaded with off topic content"


Advertising


Advertising
I NEED HELP WITHSONY VEGAS PRO 8!! Ugh and ihave a
video project due tomorrow for merrilllynch :(( all ineed to
do is simple: Extract several scenes from a clip, insert
captions, transitions and thatsit. really. omggicant ﬁgure out
anything!! help!! and igot food poisoning from eggs. its not
fun. Pleasssse, help? :(

Learning from Multi-topic Web Documents for Contextual
Advertisement, Zhang, Y., Surendran, A. C., Platt, J. C., and
Narasimhan, M.,KDD 2008


Preliminary Results in…

Identifying intents behind user posts on social
networks

‣ Identify Content with monetization potential
Identifying keywords for advertising in user-
generated content

‣ Considering interpersonal communication & off-topic
chatter


Investigations

User studies

‣ Hard to compare activity based ads to s.o.t.a
‣ Impressions to Clickthroughs
‣ How well are we able to identify monetizable posts
‣ How targeted are ads generated using our " keywords
vs. entire user generated content

Identifying Monetizable Intents
Scribe Intent not same as Web Search Intent 1B.
People write sentences, not keywords or phrases
Presence of a keyword does not imply
navigational / transactional intents

‣ ‘am thinking of getting X’ (transactional)
‣ ‘I like my new X’ (information sharing)
‣ ‘what do you think about X’ (information seeking)
1B. J. Jansen, D. L. Booth, and A. Spink, “Determining the informational, navigational, and transactional intent of web
queries,”Inf. Process. Manage., vol. 44, no. 3, 2008.


From X to Action PaTerns

Action patterns surrounding an entity

‣ How questions are asked and not topic words that indicate
what the question is about
‣ “where can I ﬁnd a chottopspcam”
‣ User post also has an entity


Conceptual Overview
Bootstrapping to learn IS paTerns
Set of user posts from SNSs
Not annotated for presence or absence of any intent


Bootstrapping to
learn IS paTerns
Generate a universal set of n-‐‑ gram paMerns; freq > f
S = set of all 4-‐‑grams; freq > 3


Bootstrapping to
learn IS paTerns
! !
Generate set of candidate paMerns from seed words
(why,when,where,how,what)

Sc= all 4-‐‑grams in S that extract seed words


Bootstrapping to
learn IS paTerns

! !
User picks 10 seed paMerns from Sc

Sis= ‘does anyone know how’, ‘where do I ﬁnd’,
‘someone tell me where’…


Bootstrapping to
learn IS paTerns
! !
! !

Gradually expand Sis by adding Information
Seeking paDerns from Sc


Bootstrapping to
learn IS paTerns
! !
! !

For every pis in Sis generate set of ﬁller paMerns


Bootstrapping to
learn IS paTerns
‘.* anyone know how’‘
does .* know how’

‘does anyone .* how’ ‘does anyone
know .*’


Extracting and Scoring PaTerns



•‘does * know how’
–‘does someone know how’
•Functional Compatibility -‐‑Impersonal pronouns
•Empirical Support –1/3

–‘does somebody know how’
•Functional Compatibility -‐‑Impersonal pronouns
•Empirical Support –0
•PaMern Retained

–‘does john know how’
•PaMern discarded


Sc= {‘does anyone know how’, ‘where do I ﬁnd’,
‘someone tell me where’}
pis= `does anyone know how’



pis= `does anyone know how’


Expanding the PaTern Pool

Functional properties / communicative functions
of words
From a subset of LIWC

–cognitive mechanical (e.g., if, whether, wondering, ﬁnd)

•‘I am thinking about geMing X’

–adverbs(e.g., how, somehow, where)

– (e.g., someone, anybody, whichever)

•‘Someone tell me where can I ﬁnd X’

1Linguistic Inquiry Word Count, LIWC, hMp://liwc.net


Details in [WISE2009] for..

Over iterations, single-‐‑word substitutions,
functional usage and empirical support
conservatively expands Sis

Infusing new paMerns and seed words
Stopping conditions


Sample Extracted PaTerns


Identifying Monetizable Posts

Information Seeking paMerns generated oﬄine
Information seeking intent score of a post
‣ Extract and compare paMerns in posts with
extracted paMerns
‣ Transactional intent score of a post
‣ LIWC ‘Money’ dictionary -‐‑ 173 words and
word forms indicative of transactions, e.g.,
trade, deal, buy, sell, worth, price etc.


Keywords for Advertizing

Identifying keywords in monetizable posts
" –Plethora of work in this space
Off-topic noise removal is our focus
" I NEED HELP WITHSONY VEGAS PRO 8!! Ugh
and ihave a video project due tomorrow for
merrilllynch :(( all ineed to do is simple: Extract
several scenes from a clip, insert captions,

Keywords for Advertising

Identifying keywords in monetizable posts

‣ Plethora of work in this space
Off-topic noise removal is our focus

‣ I NEED HELP WITHSONY VEGAS PRO 8!! Ugh and
ihave a video project due tomorrow for merrilllynch :(( all
ineed to do is simple: Extract several scenes from a clip,
insert captions, transitions and thatsit. really. omggicant
ﬁgure out anything!! help!! and igot food poisoning from
eggs. its not fun. Pleasssse, help? :(


Conceptual Overview
(also see slides 88,89)
Topical hints

‣ C1 -['camcorder']
Keywords in post

‣ C2 -['electronics forum', 'hd', 'camcorder', 'somethin', 'ive', 'canon',
'little camera', 'canon hv20', 'cameras', 'offtopic']

Move strongly related keywords from C2 to C1 one-by-one

‣ Relatedness determined using information gain
‣ Using the Web as a corpus, domain independent


Oﬀ-‐‑topic ChaTer

C1 -['camcorder']
C2 -['electronics forum', 'hd', 'camcorder', 'somethin',
'ive', 'canon', 'little camera', 'canon hv20', 'cameras',
'offtopic']
Informative words

‣ ['camcorder', 'canon hv20', 'little camera', 'hd', 'cameras',
'canon']


Evaluations -‐‑User Study

Keywords from 60 monetizable user posts

‣ Monetizable intent, at least 3 keywords in content
45 MySpace Forums, 15 Facebook Marketplace, 30
graduate students

‣ 10 sets of 6 posts each
‣ Each set evaluated by 3 randomly selected users
Monetizable intents?

‣ All 60 posts voted as unambiguously information seeking in intent


1. Eﬀectiveness of using
topical keywords
Google AdSenseads for user post vs. extracted
topical keywords


Instructions –User Study


Result -‐‑2X Relevant Impressions

Users picked ads relevant to the post

‣ At least 50% inter-evaluator agreement
For the 60 posts

‣ Total of 144 ad impressions
‣ 17% of ads picked as relevant
For the topical keywords

‣ 40% of ads picked as relevant

2. Proﬁle Ads vs. Activity Ads

User’s proﬁle information

‣ Interests, hobbies, TV shows..
‣ Non-demographic information
Submit a post
Looking to buy and why (induced noise)
Ads that generate interest, captured attention


Result -‐‑8X Generated Interest
Using proﬁle ads

‣ 7% of ads generated interest
Using authored posts

•" Using topical keywords from authored posts



To note…

User studies small and preliminary, clearly suggest

‣ Monetization potential in user activity
‣ Improvement for Ad programs in terms of relevant
impressions
Evaluations based on forum, marketplace

‣ Verbose content
‣ Status updates, notes, community and event
memberships…
‣ One size may not ﬁt all

To note…

A world between relevant impressions and click
throughs

‣ Objectionable content, vocabulary impedance, Ad
placement, network behavior
In a pipeline of other community efforts
No proﬁle information taken into account
Cannot custom send information to Google AdSense


SENTIMENT / OPINION
MINING


Content Analysis: Sentiment
Analysis/Opinion Mining
Two main types of information we can learn from
user-generated content: fact vs. opinion
Much of what we read in social media (e.g., blogs,
Twitter, Facebook) is a mix of facts and opinions.
For example, " Latest news: Mobile web services not
working in #Bahrain and Internet is extremely slow
#feb14 {fact}... looks like they "learned" from #Egypt
{opinion}"


Sentiment Analysis Motivation

Why do
Which movie What customers people oppose
should I see? complain about? health care
reform?


Sentiment Analysis: Tasks

Example:

‣ How awful that many #Egyptian artifacts are in danger of
being destroyed.
‣ What Zahi Hawass must be thinking #jan25 (read in the
tone of “what were YOU thinking”



Classiﬁcation: overall sentiment polarity: positive/
neutral/negative
‣Example: “How awful that many #Egyptian artifacts are
in danger of being destroyed.”
‣overall polarity is negative
‣Target-speciﬁc sentiment polarity: positive/neutral/
negative
‣ Example: for target "egyptian artifacts", polarity is
"negative“ for target "Zahi Hawass", polarity is "neutral“



Identiﬁcation & Extraction: opinion, opinion
holder, opinion target
Example: opinion="awful", opinion holder="the
author", target="egyptian artifacts are in danger"
Opinion="must be thinking", opinion holder="the
author", target="Zahi Hawass"


Sentiment Analysis: Approaches

Classiﬁcation:

‣ Supervised:
‣ labeled training data
‣ features, differ from traditional topic classiﬁcation tasks
‣ learning strategies
‣ Unsupervised:
‣ lexicon-based approach
‣ Bootstrapping


Identiﬁcation & Extraction:
‣utilizing the relations between opinion and opinion target,
‣proximity,
‣syntactic dependency,
‣co-occurrence and
‣prepared patterns/rules


Sentiment Analysis:
From Tweets to polls
corpus:
• 0.7 billion tweets,
Jan 2008 – Oct
2009
Jan 2008 – May
2010

Lexicon-based approach for sentiment analysis of tweets:
subjective lexicon from OpinionFinder (Wilson et al., 2005)
Within topic tweets, count messages containing these positive and
negative words deﬁned by the lexicon


Sentiment Analysis:
corpus:
Jan 2008 – Oct
2009
Jan 2008 – May
2010

subjective lexicon from OpinionFinder (Wilson et al., 2005)


Sentiment Analysis:
corpus:
Jan 2008 – Oct
2009
Jan 2008 – May
2010



Sentiment Analysis:
corpus:
Jan 2008 – Oct
2009
Jan 2008 – May
2010
B.O’Connor, R.Balasubramanyan, B.R.Routledge, and
N.A.Smith. From Tweets to polls: Linking text sentiment to public
opinion time series. In Intl.AAAI Conference on Weblogs and
Social Media, Washington,D.C.,2010.


Sentiment Analysis: Predicting
the Future With Social Media

Corpus: 2.89 million tweets referring to 24 movies released over a period of three months
Sentiment Analysis Classifier:
DynamicLMClassifier provided by LingPipe linguistic analysis package
thousands of workers from the Amazon Mechanical Turk to assign
sentiments (positive, negative, neutral) for a large random sample of tweets
train the classifier using an n-gram model
S. Asur and B.Huberman. Predicting the Future With Social Media. 2010. hMp://arxiv.org/abs/1003.5699



Sentiment Analysis Classiﬁer:


Sentiment Analysis: Target-‐‑specific opinion
identification & Classification of
Tweets-‐‑Unsupervised Approach
Simple lexicon-‐‑based method doesn'ʹt work.

Observations:
The opinions may not contribute toward the given target (1,2,3,6)
The subjectivity and polarity of opinion clues are domain-
dependent (5,7)
Single words are not enough (4,7,8)

General subjective lexicon
‣ Commonly used subjective lexicon + popular slangs learned from
Urban Dictionary
Domain-‐‑dependent sentiment lexicon
‣ Learned from domain-‐‑speciﬁc corpus
‣ bootstrapping
‣ More than words (word/phrase/paMern)
‣ n-‐‑gram + statistical model



Domain-‐‑dependent sentiment lexicon
‣ Learned from domain-‐‑speciﬁc corpus
‣ bootstrapping
‣ More than words (word/phrase/paMern)
‣ n-‐‑gram + statistical model


Sentiment Analysis: Target-‐‑
specific opinion identification &
Classification of Tweets-‐‑
Unsupervised Approach


Sentiment Analysis: Target-‐‑
specific opinion identification &
Classification of Tweets-‐‑
Unsupervised Approach
Target-‐‑specific opinion identification/extraction
‣ Shallow syntactic analysis
‣ Rules + Proximity


Content Analysis: Context
Extraction, Utilization
URL Extraction is for Tweets
FourSquare in Facebook, TwiMer
What is it in other mediums/SMS?


Content Analysis:
URL extraction
Resolution
Semantic Context Relevance


Author Categorization: Using
Content to derive additional
People metadata
Personality Signals
Blogs, Style of Writing
Psychometric analysis of content
Sample study: Gendered writing styles online


People Analysis: Using Network
to derive People metadata
Interesting questions to ask:

‣ Who are the most popular people* in the network
‣ Who are the most inﬂuential people in the network
‣ Who are the most active people in the network
‣ What are the types of people in communities of the
network
‣ Who are the bridges between communities in the network


People Analysis: Inﬂuence

By Link Analysis Algorithms
Hits [K-99] & variants
PageRank [BP-97] & variants etc..
Links not sufﬁcient!

‣ Million Follower Fallacy [C-10]

Source : informing-arts



Flavor of Context Analysis (activity level)
Popularity NOT = Influence!
‣ Influence & Passivity [RGAH-10]
Interest Similarity
‣ TwitterRank: Reciprocity & Homophily [WLJH-10]
Klout Score - True Reach, Amplification [Klout]


People Analysis: User types
& Aﬃliation
Blogger, Scientist, Journalist, Artist, Trustee,
Company X in Domain Y..

‣ Multiple types and afﬁliations!
User interest mining

‣ Key Phrase Extraction followed by semantic association on
user bio, tweets, lists, favorite posts Source: kahunainstitute.com

‣ Twitter Study [BCDMJNRM-09]


& Aﬃliation


& Aﬃliation
Semantic analysis of proﬁle description
‣ Web Presence: Use of Web & Knowledge bases
(Wikipedia, Blogs) to build context for user types
‣ Entity Spotting & Extraction, followed by Semantic
Association and Similarity with user-type context


People Analysis:
Social Engagement

Source: http://www.syscomminternational.com/

Frequency Distribution Analysis of user activity
‣ posting, retweet, reply, mentions, lists etc.


Network Analysis

Foundation of network:
•Nodes
•Connections/Relationships

Interesting questions to ask:

How communities form around topics-‐‑ growth & evolution

What are the effects of presence of influential participants in the
communities

What are the effects of content nature (or sentiment, opinions)
flowing in network on the community life

What is the community structure: degree of separation and sub-‐‑
communities

Network Analysis: Methods

Source: http://www.kudos-
dynamics.com/



Network Structure metrics
Centrality, Connected Component, Avg.
Degree, Clustering Coeﬃcient, Avg. Path Length,
Bridge, Cohesion, Prestige, Reciprocity
Important Literature:
[AB-‐‑02, WS-‐‑98, BW-‐‑00; NW-‐‑06, WF-‐‑92, MW-‐‑10]
dynamics.com/


Network Analysis: Algorithms

Community Discovery, growth, evolution

‣ Based on relationship types (e.g., signed network),
geography/location based etc.
Hierarchical clustering algorithms – Top-down,
bottom-up
Modularity Maximization [NW-06]
Algorithms comparison survey [B-06]


Network Analysis: Algorithms

Graph Partitioning & Traversal
Best time-complexity & reachability
Follow Greedy paths

‣ K-way multilevel Partitioning ,
‣ Bron-Kerbosch, K-plex, K-core or N-cliques, DFS, BFS,
MST
"ʺWe dream in Graph and
We analyze in Matrix”-‐‑
Barry Wellman, INSNA



Network Modeling Approaches

‣ Random graph model (Erdos-Renyi model)
‣ Small-world model (Small World Phenomenon)
‣ Scale-free model (led to Power-Law degree distribution)
‣ Social Network Analysis methods
‣ Centrality (Degree, Eigenvector, Betweenness, Closeness)
‣ Clusters (Cliques and extensions, Communities)

dynamics.com/


Network Analysis:
Diﬀusion & Homophily
Information Flow: Diffusion

‣ Maximizing Spread (Opinion, Innovation, Recommendation)
‣ Outbreak Detection (e.g., disease)
Social Network: No info about user action–
Understanding dynamics is challenging!
Power Law distribution [LAH-07]
Factors impacting ﬂow:

‣ Sampling strategy, user Homophily, content nature
[CLSCK-10, NPS-10]

Querying


Analysis & Visualization Tools

(Network WorkBench)NWB

Truthy

Graph-tool
Orange

Pajek
Source: hMp://truthy.indiana.edu/

Tulip

http://en.wikipedia.org/wiki/
social_network_analysis_software


Event Detection


Citizen Sensing in Real-‐‑time


Real-‐‑Time Motivation
People cant wait for Information
500 years ago

‣ Single life time

20 years ago

‣ Next day or two

‣ Television, News papers

Presently

‣ Minutes are not considered fast enough

‣ Digital media, Social media


Real-‐‑Time Social Media

Is Real-Time the future of Web?
Social Media for Real-Time Web

‣ Disaster Management
‣ Ushahidi
‣ Real-Time Markets
‣ Examples
‣ Brand Tracking
‣ Twarql
‣ Movie reviews

Scenario

The
Guardian
Feb
2010


Scenario

The
Guardian
Feb
2010

Journalist


Challenges

Information Overload

‣ Can we aggregate, organize and collectively analyze data

Real Time

‣ Can we deliver the data as it is generated


A Semantic Web Approach
Expressive description of Information need

‣ Using SPARQL (Instead of traditional keyword search)
Flexibility on the point of view

‣ Ability to "slice and dice" the data in several dimensions: thematic,
spatial, temporal, sentiment etc..

Streaming data with Background Knowledge

‣ Enables automatic evolution and serendipity
Scalable Real-Time delivery

‣ Using sparqlPuSH (SFSW'10)


Concept Feed


Architecture


Social Sensor Server


Metadata Extractions
(Social Sensor Server)
Named Entity Recognition

‣ 2 Million Entities from DBPedia
‣ Load as Trie for efﬁciency
‣ N-grams matched
‣ Example: Obama, Barack Obama


URL, HashTag Extraction

‣ Regex extraction
‣ Resolution
‣ URL Resolution: Follows http redirects for resolution
‣ HashTag Resolution: Tagdef, Tagal,WTHashTag.com


Other Metadata provided by Twitter
‣ User proﬁle: User Name, Location, Time etc..
‣ Tweet: RT, reply etc..


Structured Data
RDF Annotation

‣ Common RDF/OWL Vocabularies
‣ FOAF - (foaf-project.org) Friend of a Friend
‣ SIOC - (sioc-project.org) Semantically Interlinked
Online Communities

‣ OPO - (online-presence.net) Online Presence Ontology
‣ MOAT - (moat-project.org) — Meaning Of A Tag


Structured Data


Structured Data

A snippet of the annotation
<http://twitter.com/ bob/statuses/123456789>
rdf:type   sioct:MicroblogPost ;
sioc:content ”Fingers crossed for the upcoming #hcrvote”
sioc:hascreator   <http://twitter.com/bob> ;
foaf:maker    <http://example.org/bob> ;
moat:taggedWith   dbpedia:Healthcare_reform .
<http://twitter.com/bob> geonames:locatedIn
Dbpedia:Ohio .


Semantic Publisher


Semantic Publisher

Virtuoso to store triples
Queries formulated by the users are stored
SPARQL protocol over the HTTP to access rdf from
the store
Combine data from tweet with the background
knowledge in the rdf store


Application Server & Distribution
Hub


Application Server & Distribution
Hub
Distribution Hub
‣ PUSH Model -‐‑ Pubsubhubbub protocol
‣ Pushes the tweets to the Application Server

Application Server
‣ Delivers data to the Clients
‣ RSS Enable Concept feeds


Brand Tracking -‐‑ Example
Background Knowledge (e.g. DBpedia)

@anonymized
Lorem ipsum bla bla this is an example tweet

?category
skos:subject

? skos:subject
competitor skos:subject

moat:taggedWith

dbpedia:IPad
?tweet



?category
skos:subject

? skos:subject
competitor bla this is an example tweet
@anonymized skos:subject
Lorem ipsum bla

moat:taggedWith

dbpedia:IPad
?tweet



category:Wi-Fi
category:Touchscreen

?category
skos:subject

? skos:subject
Lorem ipsum bla

moat:taggedWith

dbpedia:IPad
?tweet


IPhone
HPTabletPC
category:Wi-Fi
category:Touchscreen

?category
skos:subject

? skos:subject
Lorem ipsum bla

moat:taggedWith

dbpedia:IPad
?tweet


1242 Articles from Nytimes
Around 800,000 tweets


President Obama
1242 Articles from Nytimes lays out plan for
Around 800,000 tweets Health care reform
in Speech to Joint
Session of Congress
(10th Sept
Timeline.com)


President Obama
1242 Articles from Nytimes lays out plan for
Around 800,000 tweets Health care reform
in Speech to Joint
Session of Congress
(10th Sept
Timeline.com)
Obama taking an
active role in Health
talks in pursuing his
proposed overhaul
of health care
system. (13th Aug

Twarql on Linked Open Data


Emerging Research Areas


Spam in Social Networks

Reasons for spamming include:

‣ Gaining Popularity
‣ Use of popular topic related keywords (e.g. hashtags of
trending topics) to propagate something off topic.

Launching malicious attacks

‣ Phishing attacks, virus, malware etc.
‣ Misleading the masses
‣ Propagating false information [MM-10].


Gaining popularity using trending keywords:
This tweet uses #Cairo but refers to a fashion
website.



Gaining popularity using trending keywords:
This tweet uses #Cairo but refers to a fashion
website. Egypt
Protests



Spam detection

‣ Content-based features
‣ Content Size, URL type, spam words
‣ Metadata-based features
‣ Account information, behavior.
‣ Network-based features
‣ Provenance. (e.g. content from a reliable source)


Trust in Social Networks

Reputation, Policy, Evidence, and Provenance used
to derive trustworthiness.
Illustrative examples of online cues used for trust
assessment.

‣ Wikipedia: article size, number of references, author, edit
history, age of the article, edit frequency etc.
‣ Product Reviews: number of helpful, very helpful ratings,
author expertise, sentiments in comments received for a
review etc.



We propose trust ontology[AHTS-10] that

‣ Captures semantics of trust.
‣ Enables representation and reasoning with trust.
Semantics of Trust speciﬁes, for a given trustor and
trustee, the following features.

‣ Type - Type of trust relationship.
‣ Scope - Context of the trust relationship.
‣ Value - Quantiﬁes the trust relationship.



Gleaning primitive (edge) trust

‣ Trust value between two nodes is quantiﬁed using
numbers. E.g., [0,1] or [-1,1] or partial ordering[TAHS-09].
Gleaning composite (path) trust

‣ Propagation via chaining and aggregation (transitivity)
Some popular algorithms for trust computation

‣ Eigentrust, Spreading Activation, SUNNY etc.


Integrating Social And
Sensor Networks
Machine sensor observations are quantitative in
nature, while human observations can be both
qualitative and quantitative.
Beneﬁts of combining observations from humans
and machine sensors

‣ Complementary evidence.
‣ Corroborative evidence


Integrating Social And
Sensor Networks
Applications of integrating heterogeneous sensor
observations

‣ Situation Awareness by using human observations to
interpret machine sensor observations.
‣ Enhancing trustworthiness using corroborative evidence.


Mobile Social Computing

Instant Discovery: Geo-‐‑tagging and location-‐‑
aware services, in combination with search, have
made discovery a two-‐‑way street.
Compressed Expression: Mobile makes social
networking even more compelling
Outsourced Memory: Cloud-‐‑based servers to
store all of their mobile applications and
databases



Compressed Expression: Mobile makes social
networking even more compelling
databases



databases



Automated Decisions: Smart apps helps to make
faster decisions or even apps makes decisions for
us
Peer Power: Mobiles can create social movements
based on peer inﬂuence


Mobile Social Computing (Cont.)

Personalized Branding: advertising are rapidly
becoming personalized based on individual's needs
and preferences
Mobiles in social development becoming an integral
part of development

‣ Coordination in disaster situations
‣ Health care delivery, especially in developing countries
‣ Elections and other forms of political expression


Research Application: Twitris


Twitris -‐‑ Motivation

1. Information Overload
Multiple events around us
WHAT to be aware of
Multiple Storylines about same event!!



2. Evolution of Citizen Observation

‣ with location and time



3. Semantics of Social perceptions

‣ What is being said about an event (theme)
‣ where (spatial)
‣ When (temporal )
Twitris lets you browse citizen reports using social
perceptions as the fulcrum


Twitris: Semantic Social Web
Mash-‐‑up
Facilitates understanding of multi-‐‑dimensional social perceptions over
SMS, Tweets, multimedia Web content, electronic news media


Twitris: Architecture


Citizen Sensing, Social Media Analytics, and Applications

Citizen Sensing, Social Media Analytics, and Applications

Recommended

Recommended

More Related Content

What's hot

What's hot (7)

Similar to Citizen Sensing, Social Media Analytics, and Applications

Similar to Citizen Sensing, Social Media Analytics, and Applications (20)

Recently uploaded

Recently uploaded (20)

Citizen Sensing, Social Media Analytics, and Applications