Le  nuove  tecnologie  di  Social  Networking  e  le  imprese  

Giuseppe  Manco  ICAR-­‐CNR  

manco@icar.cnr.it    

Giuseppe  Manco  

•  Ricercatore  presso  ICAR-­‐CNR  •  Aree  di  interesse  –  Data  Analysis,  Social  Networks,  sistemi  di  recommendaCon  

CONNETTERSI,  COMUNICARE  

Sei  gradi  di  separazione  

Myspace:  110  milioni  di  uten?  

Cos’è  una  rete  sociale?  Traditional Media

Broadcast Media: One-to-ManyBroadcast Media: One to Many

Communication Media: One-to-One

Cos’è  una  rete  sociale?  

•  User  generated  content  •  User  enriched  content  •  User  interacCon  •  Comunicazione  ubiqua  •  CollaboraCve  environment  •  …  

Characteristics of Social Media� Everyone can be a media outlet� Disappearing of communications barrier

� Rich User Interaction� User-Generated Contents� User Enriched Contents

User developed widgets� User developed widgets� Collaborative environment� Collective WisdomC� Long Tail

Broadcast MediaFilter, then Publish

Social MediaPublish, then Filter

User  generated  content  

User  interac?on  

Comunicazione  ubiqua  

AEvità  condivise  

I  websites  più  visitaC  

•  Il  traffico  internet  più  alto  (daC  Alexa,  oQobre  2012)  

1   Facebook   11   Blogspot  

2   Google   12   LinkedIn  

3   YouTube   13   Taobao  

4   Yahoo   14   Google  India  

5     Baidoo   15   Yahoo  Japan  

6   wikipedia   16   Sina.com.cn  

7   Windows  live   17   msn  

8   TwiQer   18   Google  hk  

9   QQ.com   19   Google  de  

10   Amazon   20   Bing  

Il  ruolo  fondamentale  dei  social  media  

TIPOLOGIE  DI  RETE  

L’Universo  Social  Media  

Social  Media  

Social  Networking  

Blogs  

Wki  Forum  

Content  sharing  

QUALI  OPPORTUNITÀ?  

L’azienda  e  i  social  networks  •  Pubbliche  relazioni  •  Customer  Support  •  Market  Research  •  Brand  MarkeCng  •  PromoCons  •  Consumer  EducaCon  •  Sales  •  New  Product  Development  •  Customer  RelaConship  Management  

SOCIAL  MEDIA  ANALYTICS  

DaC,  daC,  daC  •  Grossi  volumi,  grande  varietà  

–  Milioni  di  utenC,  milioni  di  contenuC  –  testuale,  MulCmediale  (immagini,  

video,  etc.)    –  Milioni  di  connessioni  –  Tendenze,  preferenze,  

comportamenC,  …  •  I  daC  sono  open  e  facili  da  accedere  

–  Facili  da  reperire  –  Di  pubblico  dominio  –   Developers  APIs  –   Spidering  the  Web  

Le  opportunità  

•  Ogni  utente  può  condividere  e  contribuire  ai  contenuC,  esprimere  opinioni,  collegarsi  ad  altri  

•  Questo  significa:  – Human  behavior  – MarkeCng  analyCcs    – Product  senCment  

� Any user can share and contribute content, express opinions, link to others

� This means: Can data-mine opinions and behaviors of millions of users to gain insights into: � Human behavior � Marketing analytics � Product sentiment

Jure Leskovec:Social Media Analytics (KDD '11 tutorial) 6 8/21/2011 6

Actionable Intelligence

Consumer Generated, Not Edited, Not Authenticated

8/21/2011 Jure Leskovec:Social Media Analytics (KDD '11 tutorial) 7

Applicazioni:  ReputaCon  management  

•  Consumer  Brand  AnalyCcs  –  Cosa  dice  la  gente  sul  mio  marchio?    

•  MarkeCng  CommunicaCons  – Determinare  se  le  campagne  che  pianifico  saranno  efficaci  

•  Product  reviews  –  Estrazione  automaCca  di  review  e  informazioni  su  prodom  e  servizi  •  Facile  da  usare,  confortevole,  prezzo  adeguato,  …    

Applicazioni:  Responsività  

•  CiCzen  response    •  feedbacks  su  temaCche  poliCche  •  Campagne  poliCche –  Perché  la  gente  supporta  un  candidato?    

•  Law  enforcement  – MovimenC  dissidenC  su  TwiQer    – Minority  report    hQp://www.nyCmes.com/2011/08/16/us/16police.html?_r=1    

Applicazioni:  Viral  MarkeCng  •  Viral  markeCng:    –  Raccomandazioni  personlizzate  

•   Il  ruolo  dei  forum  online:  –  79.2%  dei  partecipanC  ai  forum  aiutano  gli  utenC  connessi  a  prendere  decisioni  relaCve  a  un  prodoQo  

–  65%  dei  partecipanC  ai  forum  condividono  consigli  (offline  o  personalizzaC)  basaC  sulle  informazioni  che  hanno  leQo  online  

     hQp://www.socialmediaexaminer.com/new-­‐studies-­‐show-­‐value-­‐of-­‐social-­‐me  

Applicazioni:  Human  Behavior  analysis  

•  Processare  I  contenuC,  e  usufruire  di  tools  per    –  IdenCficare  reC  sociali:  gruppi,  membri  –  IdenCficare  topics  e  senCments  

� Process social media content, provide tools for analysts to: � Identify social networks: groups, members

� Identify topics and sentiment

Social Media Content

Link Diagrams

Predictive Modeling

8/21/2011 Jure Leskovec:Social Media Analytics (KDD '11 tutorial) 12

Relevance,  Authority,  SenCment  

•  Le  tre  dimensioni  dell’interazione  sociale  

Finding the Relevant BlogsOUR FIRST OBJECTIVE is to filter the vast blogosphere

from millions to the thousands of blogs most relevant to thetopic of interest. In the simplest case, the “topic” can be a spe-cific product, and the objective is therefore to identify all blogsdiscussing this product and perhaps competing products aswell. More generally, one would like to cast a wider net andinclude blogs that are discussing higher-level concepts relatedto the market addressed by the product(s) of interest. Forexample, IBM offers a social networking product called LotusConnections, but marketing experts may wish to follow all dis-cussions touching on the concept of collaborative software as ameans to understand emerging trends in this space.

The distinction between tracking a specific product andtracking a broader concept impacts the methodology used tofind the relevant set of blogs. If the interest is only in a specificproduct, it is straightforward to identify blogs (e.g. by using ablog search engine) containing references to the product. Suchan approach is less effective for broad topics because discus-sions that touch on such a topic (e.g.“collaborative software”)may not specifically contain these keywords. In practice, it isreasonable to ask marketing experts to identify a small set of“seed” blogs that are highly relevant to the topic at hand. Oneapproach is to use these labeled blogs to build a straightforwardtext classification model to identify other relevant blogs.

Relevant blogs are likely to link to other relevant blogs, and analternative approach to text classification is to exploit the structureof the blog cross-reference graph. One simple approach is to startwith the small set of expert-identified seed blogs,add all the blogsthey link to and then repeat this process for several iterations(degrees of separation). This snowball sampling procedure wasused to identify the blogs shown in Figure 1; note that the second(and third) iteration of this process identified a number of rele-vant blogs not included in the seed population.

Discussions about broad concepts like “collaboration soft-ware” tend to be tightly connected, and hence this simpleapproach is likely to be more efficient than keyword search infinding these blog sub-communities. Using the graph structurealso alleviates the problem when product search terms havemultiple meanings, e.g.“Lotus” is a car, a flower and a softwarebrand – it is unlikely that blogs talking about Lotus the car willreference blogs discussing Lotus Software.

An important consideration is to avoid crawling [computerprogram that browses the Internet in a methodical, automatedmanner] the parts of the relevant blog sub-universe that areirrelevant from a marketing perspective. A practical solution isfocused snowball sampling [3], which explicitly focuses Web

that utilize these analytics capabilities to pro-vide marketing insights from blogs.

Social Media Analytics for Marketing

FROM A MARKETING and marketintelligence perspective, blogs are a veryimportant form of social media becausethey provide access to previously inaccessi-ble information such as specific customer insights and opin-ions. Social media analytics can address several interestingquestions by providing algorithms and approaches for theautomated analysis of blogs and related social media:1. Given the massive size of the blogosphere, how can we

identify the subset of blogs and forums that arediscussing not only a specific product, but also higher-level concepts that are in some way relevant to thisproduct?

2. What sentiment is expressed about a product orconcept in a blog or forum?

3. Who are the most authoritative or influential bloggersin this relevant subspace?

4. What are the novel emerging topics of discussionhidden in the constant chatter in the blogosphere?

A typical blog or micro-blog has one author (the blogger)and consists of multiple entries or posts. It is useful to think ofa blog in a three-dimensional space defined by the first threemetrics above: relevance, sentiment and authority. While thefirst two dimensions, relevance and sentiment, are specific to agiven post or even smaller section of text (“snippet”), thenotion of authority is most naturally assigned at the blog level.A blogger’s authority can also depend on the specific topic.Emerging topics are a property of the blogosphere at large andrequire analysis across many blogs.

All three dimensions are important and they need to be con-sidered in a unified view in order to provide marketing insight.One way to provide such a view is to determine the relevanceand sentiment of each post, and characterize the overall rele-vance and sentiment of the blog as a simple statistic over indi-vidual posts.

Figure 1 captures such a blog-centric view along thesedimensions from a prototype tool at IBM Research. Here, weare interested in a high-level view of blogs relevant to the broadtopic of “social collaboration.” Relevance is shown on the y-axis, and sentiment is on the x-axis – both metrics are com-puted at the post level and aggregated to the blog level. Eachcircle represents a blog, and the size of the circle reflects theblogger’s authority. The output of the model can be interpret-ed as the probability that the post is positive in tone. We aremost interested in extremes of sentiment, so we naturally lookfor authoritative blogs in the upper-left and upper-right quad-rants to find the most relevant blogs with non-neutral senti-ment. Such a view allows marketing people to quickly identifyblogs of interest, and to drill down to obtain more specificunderstanding of the potential marketing impact.

www.orms-today.com 27

Figure 1: Relevance, authority and sentiment at the blog level.

ORMS3701_FTRs 2/3/10 4:56 PM Page 27

IBM’s  topic-­‐based  blog  evaluator    

SenCment  DetecCon  

•  E’  possibile  caraQerizzare  (in  maniera  automaCca)  il  tono  di  una  discussione?  

crawling using only links deemed to be relevant by a text classi-fier. There are many other opportunities to apply large-scale ver-sions of classification models [4] that exploit both graph struc-ture and text content.

Sentiment DetectionWHILE THE RELEVANCE MODEL can help to limit the

universe to thousands rather than a half million blogs, this is still avolume far beyond close scrutiny by a small marketing staff. Butwhich ones of all the relevant blogs should you read? And are themost relevant indeed the most crucial? An immediate concernfrom a marketing perspective is to detect strong sentiment,partic-ularly strong negative sentiment. Since it is impossible to read allblog posts relevant to a particular topic, there is strong motivationto develop automated capabilities to characterize the tone andsentiment in these discussions.

Sentiment detection models are crucial in order to identifyblog posts that may require swift marketing action. Figure 2 illus-trates such a situation identified by an IBM social media analyticstool. Sentiment models are able to detect a negative post (1),resulting in a rapid response (2) from a product executive. Thequick response leads to a very positive statement (3) from theoriginal blogger. This exchange illustrates the role social mediaanalytics can play in allowing marketing to identify and addressnegative sentiment before it can cause more brand damage.

The main challenge in sentiment classification is that theexpression of sentiment tends to be domain specific, and the setof domains to monitor change often. Thus we require sentimentclassifiers that can rapidly adapt to new domains withoutrequiring a large number of manually labeled training examplesof positive and negative sentiment. Treating sentiment detectionas a text classification task has made it possible to adapt to newdomains, provided there are enough training examples in thetarget domain. However, supervision for a sentiment classifiercan be provided not only by labeling documents (e.g. blogposts), but also by labeling words. For instance, labeling a wordsuch as “atrocious” as negative is one way to express our prior

belief of the sentimentassociated with it. It is pos-sible to learn from suchlabeled words in conjunc-tion with labeled docu-ments. Furthermore, theselection of words and doc-uments to be labeled can bemade algorithmically.

Such an approach isknown as active dual super-vision [5], and it can greatlyreduce the effort required tolabel examples in a newdomain. Even though thereare expressions of sentimentthat are domain-specific,

there is still a large amount ofoverlap in how positive and

negative emotion is conveyed across domains. This enables theuse of transfer learning to adapt a classifier trained in one domainto a new domain with little to no labeled data in the targetdomain [6].

Measuring Influence and AuthorityWHILE RELEVANCE AND SENTIMENT provide two

essential filters, it is unlikely that each and every relevant blog withnegative sentiment warrants an action.An important considera-tion is how much does the opinion of one blogger actually mat-ter? A well-known riddle asks,“If a tree falls in a forest and no oneis around to hear it, does it make a sound?”This ultimately trans-lates into the question of how influential is the blog in question –is anybody actually listening (or reading) and is it likely that theseopinions will influence other individuals?

Influential bloggers may or may not be factual experts butnevertheless influence the opinions of others via discussions onthe topic. From a marketing perspective, it is important to iden-tify this set of bloggers, since any negative sentiment theyexpress could spread far and wide. In addition to authorities,there are bloggers who are very well connected, who are mostresponsible for the spread of information in the blogosphere.When presented with a large number of posts relevant to atopic, ordering them by the blogger’s influence assists in infor-mation triage, given that it is not feasible to read all posts. Figure3 shows such a view, where we have found the most authorita-tive blogs relevant to the topic “social collaboration.”

Since reliable blog readership information is difficult to obtain,the links between blogs are commonly used instead to determinea blog’s authority. For instance, Technorati (www.technorati.com/) assigns an authority score to a blog based on the numberof blogs linking to the Web site in the last six months. Similarly,Blogpulse (www.blogpulse.com/) ranks blogs based on the num-ber of times it is cited by other bloggers over the last 30 days.Given that we have a network of directed edges indicating thelinks between posts/blogs, we can apply more complex measuresof prestige from social network analysis. For instance, the author-

OR/MS TODAY28 February 2010

Figure 2: Identifying and addressing negative sentiment.

M A R K E T I N G & S O C I A L M E D I A

ORMS3701_FTRs 2/3/10 4:56 PM Page 28

IBM  Social  M

edia  AnalyCcs  

Misurare  Influence  e  Authority  

•  Chi  sono  gli  utenC  suscembili?    •  Come  si  propaga  un’informazione?  •  Quando  un’opinione  è  affidabile?  

Emerging  Topics  •  Higher-­‐level  concepts  dall’informazione  che  si  distribuisce  •  Come  variano  quesC  concem?  

Most  menConed  phrases  in  the  US  presidenCal  campaign  

hQp://mem

etracker.org  

I  social  media  e  le  imprese…  

•  Due  prospemve  – Nuovi  scenari  e  modelli  di  interazione  – AnalyCcs  

•  StreQa  cooperazione  con  ricerca  e  innovazione  – Nuovi  challenges  – Opportunità  enormi