Meena Nagarajan, Amit P. Sheth KNO.E.SIS Center Wright State University M. Nagarajan, K. Baid, A. P. Sheth, and S. Wang, "Monetizing User Activity on Social Networks - Challenges and Experiences“, 2009 IEEE/WIC/ACM International Conference on Web Intelligence, Milan, Italy

 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

May 30,June

June 01, 2009

 Interests do not translate to purchase intents  Interests are often outdated..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

June 01, 2009

CONTENT-BASED ADS ON THEIR PROFILES

 Non-trivial  Non-policed content ▪ Brand image, Unfavorable sentiments 1  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 I NEED HELP WITH SONY VEGAS PRO 8!! Ugh and i have a video project due tomorrow for merrill lynch :(( all i need to do is simple: Extract several scenes from a clip, insert captions, transitions and thats it. really. omgg i cant figure out anything!! help!! and i got food poisoning from eggs. its not fun. Pleasssse, help? :( 1 Learning from Multi-topic Web Documents for Contextual Advertisement, Zhang, Y., Surendran, A. C., Platt, J. C., and Narasimhan, M., KDD 2008

 Cultural Entities  Word Usages in self- presentation  Slang sentiments  Intentions WHAT WHY HOW

 Identifying intents behind user posts on social networks  Content with monetization potential  Identifying keywords for advertizing in user- generated content  Interpersonal communication & off-topic chatter

 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

Identification, Evaluation

 Scribe Intent not same as Web Search Intent 1  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) 1 B. 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.

 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 chotto psp cam”  User post also has an entity

Set of user posts from SNSs Not annotated for presence or absence of any intent

Generate a universal set of n-gram patterns; freq > f S = set of all 4-grams; freq > 3

Generate set of candidate patterns from seed words (why,when,where,how,what) S c = all 4-grams in S that extract seed words

User picks 10 seed patterns from S c S is = ‘does anyone know how’, ‘where do i find’, ‘someone tell me where’….

Gradually expand S is by adding Information Seeking patterns from S c S c = all 4-grams in S that extract seed words S is = ‘does anyone know how’, ‘where do i find’, ‘someone tell me where’….

For every p is in S is generate set of filler patterns S is = ‘does anyone know how’, ‘where do i find’, ‘someone tell me where’….

‘.* anyone know how’ ‘does.* know how’ ‘does anyone.* how’ ‘does anyone know.*’ Look for patterns in S c -Functional compatibility of filler -words used in similar semantic contexts - Empirical support for filler ‘does anyone know how’

 Functional properties / communicative functions of words  From a subset of LIWC 1  cognitive mechanical (e.g., if, whether, wondering, find) ▪ ‘I am thinking about getting X’  adverbs (e.g., how, somehow, where)  impersonal pronouns (e.g., someone, anybody, whichever) ▪ ‘Someone tell me where can I find X’ 1 Linguistic Inquiry Word Count,LIWC,

 S c = {‘does anyone know how’, ‘where do I find’, ‘someone tell me where’}  p is = `does anyone know how’  ‘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 ▪ Pattern Retained  ‘does john know how’ ▪ Pattern discarded

 Over iterations, single-word substitutions, functional usage and empirical support conservatively expands S is  Infusing new patterns and seed words  Stopping conditions

 doesanyoneknowhow  anyoneknowhowto  idontknowwhat  knowwhereican  tellmehowto  idontknowhow  anyoneknowwherei  doesanyoneknowwhere  doesanyoneknowwhat  anybodyknowhowto  anyoneknowhowi  imnotsurewhat  doesanybodyknowhow  doesanyoneknowwhy  iwaswonderinghow  doesanyoneknowwhen  tellmewhatto  imnotsurehow  iwaswonderingwhat  noideahowto  someonetellmehow  havenocluewhat  doesanyoneknowif  idontknowif  knowifican  anyoneknowifi  imnotsureif  iwaswonderingif  ideawhatyouare  letmeknowhow  andidontknow  nowidontknow  butidontreally  waswonderingifsomeone  wouldliketosee  seewhatican  anyonehaveanyidea  wonderingifsomeonecould  waswonderinghowi  idonotwant

 Information Seeking patterns generated offline  Information seeking intent score of a post  Extract and compare patterns in posts with extracted patterns  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.

 Training corpus  8000 user posts ▪ MySpace Computers, Electronics, Gadgets forum  309 unique new patterns, 263 unambiguous  Testing patterns for recall  ‘To buy’ Marketplace – average 81 %

Off-topic noise elimination

 Identifying keywords in monetizable posts  Plethora of work in this space  Off-topic noise removal is our focus I NEED HELP WITH SONY VEGAS PRO 8!! Ugh and i have a video project due tomorrow for merrill lynch :(( all i need to do is simple: Extract several scenes from a clip, insert captions, transitions and thats it. really. omgg i cant figure out anything!! help!! and i got food poisoning from eggs. its not fun. Pleasssse, help? :(

 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

 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']

Preliminary work

 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

 Google AdSense ads for user post vs. extracted topical keywords

 Choose relevant Ad Impressions  VW 6 disc CD changer  I need one thats compatible with a 2000 golf most are sold from years if anyone has one [or can get one] PLEASE let me know!

 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  Total of 162 ad impressions  40% of ads picked as relevant

 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

 Using profile ads  Total of 56 ad impressions  7% of ads generated interest  Using authored posts  Total of 56 ad impressions  43% of ads generated interest  Using topical keywords from authored posts  Total of 59 ad impressions  59% of ads generated interest

 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

 A world between relevant impressions and clickthroughs  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

 Keywords to Ad Impressions  Google Adsense like web service  Monetization potential of a keyword on the Web not the same on a social n/w?  Ranking keywords in user post  We are building an F8 app  Collaboration for clickthrough data

 Google/Bing: Meena Nagarajan    Google/Bing: Amit Sheth  