Uprising microblogs: A Bayesian network retrieval model for tweet search
6 likes3,658 views
The document presents a Bayesian network retrieval model for tweet search, addressing challenges in microblogging such as information overload and the need for real-time relevant data. It details the components of the model, including how to compute conditional probabilities, and evaluates its performance against traditional information retrieval models. The study concludes that the proposed model integrates various relevance factors and outperforms established baselines, suggesting paths for future enhancements in real-time tweet search capabilities.
![Microblogging service
Microblog?
“ Microblogging is a new form of communication [….]
that enables users to broadcast and share information
about their activities, opinions and status. [Java et
al.2007].
”
• Microblog post
– Short (140 characters)
1 billions Publications /week
– Real-time 50 millions Publications /day
– Social motivation 177 million Publications in mars 2011
– Mobile device +106 millions User accounts
3](https://image.slidesharecdn.com/acmsac2012-120623164834-phpapp01/85/Uprising-microblogs-A-Bayesian-network-retrieval-model-for-tweet-search-3-320.jpg)
Uprising microblogs: A Bayesian network retrieval model for tweet search
- 1. Uprising microblogs: A Bayesian network retrieval model for tweet search Lamjed Ben Jabeur, Lynda Tamine and Mohand Boughanem IRIT, Université Paul Sabatier
- 2. A Bayesian network retrieval model for tweet search Outline 1. Microblogging service 2. Tweet search 3. Bayesian network topology 4. Computing conditional probabilities 5. Experimental evaluation 6. Conclusion and future work 2
- 3. Microblogging service Microblog? “ Microblogging is a new form of communication [….] that enables users to broadcast and share information about their activities, opinions and status. [Java et al.2007]. ” • Microblog post – Short (140 characters) 1 billions Publications /week – Real-time 50 millions Publications /day – Social motivation 177 million Publications in mars 2011 – Mobile device +106 millions User accounts 3
- 4. Microblogging service Tweet, retweet et hashtag ? “ Jack Dorsey 21 Mars 06 1ier Tweet inviting coworkers #oilspill “ Stephen Colbert 21 Juin 2010 Golden Tweet Award 2010 In honor of oil-soaked birds, 'tweets' are now 'gurgles. http://bit.ly/cIhZNf “ Wendy's 8 Juin 2011 Golden Tweet Award 2011 RT for a good cause. Each Retweet sends 50¢ to help kids in foster care. #TreatItFwd “ CORIA11 16 mars 2010 CORIA 2011 : Université d'Avignon #CORIA11 http://yfrog.com/h3y ““ MohBoughanem 17 Mars 2010 @coria2011 well visualized, quickly found MohBoughanem CORIA11 17 Mars 2010 4 @coria2011 well visualized, quickly found
- 5. Microblogging service Social information network 5
- 6. Tweet search Microblog IR • Users overwhelmed by the huge quantity of tweets – Important publication rate – Diverse sources of information Difficulty to accessing to interesting posts • Microblog IR tasks – Person search and follower suggestion – Trend extraction – Opinion search – Tweet search 6
- 7. Tweet search Tweet search task “ real-time search task, where the user wishes to see the most recent but relevant information to the query. (Ounis et al., 2011). ” “ adhoc search on Twitter, where a user’s information need is ” represented by a query at a specific time. (Ounis et al., 2011). • Search motivations – access to concise and credible information – access to fresh and real-time news – follow an event – collect opinions and public sentiments 7
- 8. Tweet search Related work 1. Spatio-temporel context TwitterStand (Sankaranarayanan J. et al, 2009) TweetSieve (Grinev M et al, 2009) 2. Microblog features – followership, tweets, retweets, reply, hashtags, URLs – Linear combination (Nagmoti et al., 2010) – Learn to Rank (Duan Y et al., 2010) 8
- 9. Tweet search Related work 3. Social network structure – Indegree, Retweet et Mention influence (Cha et al., 2010).,TweetRank, FollowerRank (Nagmoti et al., 2010). – Authority (Kwak et al., 2010) – Influence (Kwak et al., 2010), TwitterRank (Weng et al., 2010), Popularity (Duan et al.,2010) 9
- 10. Tweet search Contributions topical • Relevance features: – Term occurrence – social influence – time magnitude • Bayesian network model temporal social 10
- 11. Bayesian network topology Definitions and notations • Query: q 0,1 q, q • Term: ki 0,1 k , ki i • Term configuration: k example : k1 , k 2 k k1 , k2 ), (k1 , k2 ), (k1 , k2 ), (k1 , k2 ) ( • Tweet: t j 0,1 ti , ti • Microblogger: uk 0,1 uk , uk 11
- 12. Bayesian network topology Network nodes and edges Query q Terms k1 k2 k3 Tweets t1 t2 t3 Microbloggers u1 u2 12
- 13. Computing conditional probabilities Query evaluation Query q P(q t i ) P(q | k )P(k | t i ) P( t i | u k ) P(u k ) k Terms k1 k2 k3 P(q t j ) P(q | k )P( t j | u k ) P(u k ) k Tweets t1 t2 t3 P(k i | t j ) P(k i | t j ) k |on(i,k ) 1 i k i |on(i,k ) 0 Microbloggers u1 u2 13
- 14. Computing conditional probabilities Query P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 P(q | k ) on(i, k ) i , ki q 14
- 15. Computing conditional probabilities Tweet P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 P(k i | t j ) (1 ) F (ki , t j ) H (ki , t j ) T (ki , t j ) L(t j ) Term occurrence Tweet properties P( k i | t j ) 1 P( k i | t j ) 15
- 16. Computing conditional probabilities Term frequency P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 P(k i | t j ) (1 ) F (ki , t j ) H (ki , t j ) T (ki , t j ) L(t j ) a if k i t j F ( ki , t j ) 1 1 F (ki , t j ) tf ki ,t j 0,8 0 a=0,1 otherwise 0,6 a=0,25 0,4 a=0,5 0,2 a=0,75 0 a=1 0 5 tf ki ,t j10 16
- 17. Computing conditional probabilities Hashtag P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 P(k i | t j ) (1 ) F (ki , t j ) H (ki , t j ) T (ki , t j ) L(t j ) b if # k i t j 1 H (ki , t j ) tf #ki ,t j b otherwise 17
- 18. Computing conditional probabilities Time magnitude P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 P(k i | t j ) (1 ) F (ki , t j ) H (ki , t j ) T (ki , t j ) L(t j ) tweets df k i, j T ( ki , t j ) 30 j 20 t1 10 t2 0 1 2 tems 3 4 5 j t k , t j t k t time 18
- 19. Computing conditional probabilities Tweet length P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 P(k i | t j ) (1 ) F (ki , t j ) H (ki , t j ) T (ki , t j ) L(t j ) 1 L(t j ) 1 avgtl tltj 19
- 20. Computing conditional probabilities Microblogger P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 1 P( t j | u k ) u k 20
- 21. Computing conditional probabilities Social influence P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 P(uk ) Inf (uk ) PageRank on Retweet Social Network 1 Inf G 1 (ui ) k Inf Gk (ui ) d (1 d ) w j ,i U u j ,e ( u j ,ui )E O(u j ) (u j ) (u j ) w j ,i (u j ) 21
- 22. Computing conditional probabilities Social influence P(q t j ) P(q | k )P( t j | u k ) P(u k ) P(k i | t j ) P( k i | t j ) k |on(i,k )1 k i k i |on(i,k ) 0 (u j ) (ui ) wi , j (ui ) 22
- 23. Experimental evaluation TREC 2011 Microblog NESTOR Microblog Search Engine Tweets 16 141 812 Microbloggers 5 356 432 Retweets 1 128 179 Retweet relationships 1 060 551 Tweet 1 860 112 Social network of retweets: nodes 5 495 081 Terms 7 781 775 Social network of retweets: edges 1 024 914 Hashtags 455 179 Giant component 11.12% Term frequency Hashtags Tweet length 1.5E8 1.5E 7 1.5E 6 0 5 10 0 5 10 0 20 Term frequency, hashtags and length distributions 23
- 24. Experimental evaluation Queries and ground truth • “Arab Spring” query dataset (25 queries) – Topical “Number of protesters in Tahrir”, “Tunisian revolution” – Temporal “ElBaradei arrvies in Egypt”, “Clashes in Tahrir”, “SMS Down Egypt” – Social “Wael Ghonim”, “Mubarak dissolves government” • User rating (relevant, not relevent) • Tweets ranked by Score; p@10; p@20 24
- 25. Experimental evaluation Configurations and baselines BNTS Bayesian network model for tweet search* BNTS-L BNTS, Tweet length feature disabled BNTS-T BNTS, Time magnitude feature disabled BNTS-H BNTS, Hashtag feature disabled BNTS-S BNTS, Social influence feature disabled BM25 Okapi BM25 VSM Vector Space Model BM Boolean Model * 0.25, a 0.25, b 0.4, t 1h, d 0.15 25
- 26. Experimental evaluation Features impact BNTS BNTS-L BNTS-T BNTS-H BNTS-S 0,584 0,58 0,552 0,532 0,548 0,542 0,528 0,502 0,294 0,256 p@10 p@20 26
- 27. Experimental evaluation Features impact Topical BNTS BNTS-L BNTS-T BNTS-H BNTS-S 0,7533 0,7333 0,7233 0,66 0,6867 0,6867 0,6833 0,6833 0,3767 0,2867 p@10 p@20 27
- 28. Experimental evaluation Features impact Temporal BNTS BNTS-L BNTS-T BNTS-H BNTS-S 0,4333 0,4 0,3333 0,35 0,3 0,3167 0,2333 0,2 0,1 0,0667 p@10 p@20 28
- 29. Experimental evaluation Features impact Social BNTS BNTS-L BNTS-T BNTS-H BNTS-L 0,3714 0,3286 0,3286 0,3286 0,3357 0,2714 0,2857 0,2429 0,2571 0,2 p@10 p@20 29
- 30. Experimental evaluation Retrieval effectiveness p@10 p@20 BNTS 0,552 0,548 BM25 0,576 -4% 0,494 11% BM 0,416 ** 33% 0,382 ** 34% VSM 0,376 ** 47% 0,36 ** 52% 30
- 31. A Bayesian network retrieval model for tweet search Conclusion and future work • Tweet search model – Normalized Term frequency – Time magnitude – Social influence • Integrating relevance factors within a Bayesian network • Query profile impact features performances. • Our model outperforms traditional IR baselines. • Future work – Automatically detect optimal time window – Select appropriate feature depending on the query profile 31
- 32. Thank you for your attention! Follow me on Twitter! http://twitter.com/amjedbj
- 33. Computing conditional probabilities Query evaluation q P(t j | q) P(q | k ) P(t j | k )P(k ) k k1 k2 k3 P(t j | q) P(q | k ) P(tkj | k )P(toj | k ) P(t sj | k ) P(k ) k o1 o2 u1 u1 tk1 tk2 tk3 to3 to2 to3 ts1 ts2 ts3 t1 t2 t3 33
- 34. Experimental evaluation Term frequency normalization • BNTS.K p @ 30 1 tf ki ,t j 0,35 P(t kj | k ) 0,3 k ki k t j tf ki ,t j 0,25 0,2 0,15 0,1 0,05 0 0 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 1 34
- 35. Experimental evaluation Time window • BNTS.KO p @ 30 0,32 t t oe : oe , oe 0,315 2 2 0,31 0,305 0,3 0,295 jours 0,29 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 35 t
- 36. Experimental evaluation Retrieval effectiveness isiFDL DFReeKLIM30 BNTS Médiane Nestor BM25 Disjunctive 0,5 0,45 0,4 0,35 0,3 0,25 0,2 0,15 0,1 0,05 0 p@30 MAP 36
- 37. Experimental evaluation TREC Microblogs 2011 Ranked by time Ranked by score All rel High rel All rel p@30 MAP p@30 MAP p@30 MAP Nestor* 0.2027 0.1305 0.0838 0.1287 0.2218 0.1384 Nestor-S* 0.2027 0.1305 0.0838 0.1286 0.2184 0.1360 Nestor-T 0.2082 0.1343 0.0585 0.0912 0.1912 0.1196 Nestor-L 0.2048 0.1306 0.0565 0.0867 0.2293 0.1426 Median 0.2592 0.1433 0.2646 0.1381 37
- 38. Experimental evaluation TREC Microblogs 2011 Système Seuil p@10 p@20 p@30 Map 1 Somme IDF des termes présents 30 0,3633 0,3316 0,3333 0,1759 2 BM25 30 0,3571 0,3245 0,2973 0,1546 3 Proportion des termes présents 30 0,2653 0,2561 0,2782 0,14 4 Somme des fréquences booléennes 30 0,2571 0,2663 0,2755 0,1387 5 EBM (AND) 30 0,3041 0,2918 0,2714 0,1282 6 Réseau d’inférence Bayésien 30 0,302 0,2888 0,2687 0,1274 7 Somme TF*IDF 30 0,302 0,2888 0,2687 0,1274 8 VSM 30 0,302 0,2888 0,2687 0,1274 9 Somme TF 30 0,2327 0,2276 0,2238 0,1066 10 Nestor 0,2857 0,2347 0,2027 0,1305 11 EBM (OR) 30 0,1837 0,1786 0,166 0,0541 12 Sommes des fréquences des Hashtags 30 0,1612 0,1541 0,1469 0,0512 13 Lucene-Baseline 1000 0,1612 0,1143 0,0986 0,1411 14 Somme TF (normalise par longueur) 30 0,0816 0,0673 0,0612 0,0223 15 Ordre chronologique inverse 30 0,0184 0,0255 0,0218 0,0082 38