Indian Premier League (IPL), Cricket, Online Social Media
IIndian Premier League (IPL), Cricket, Online Social Media
Megha Arora ∗ Indraprastha Institute ofInformation Technology - Delhi [email protected] Raghav Gupta * Carnegie Mellon University [email protected] PonnurangamKumaraguru
Indraprastha Institute ofInformation Technology - Delhi [email protected]
ABSTRACT
In recent past online social media has played a pivotal rolein sharing of information and opinions on real time events.Events in physical space are reflected in digital world throughonline social networks. Event based studies for such contenthave been widely done on Twitter in the computer sciencecommunity. In this report, we performed a detailed analysisof a sports event called the Indian Premier League (IPL’13)for both Facebook and Twitter. IPL is the most popularcricket league in India with players from across the world.We analysed more than 2.6 million tweets and 700 thousandFacebook posts for temporal activity, text quality, geogra-phy of users and the spot-fixing scandal which came up dur-ing the league. We were able to draw strong correlationsbetween the brand value of teams and how much they weretalked about on social media across Facebook and Twit-ter. Analysis of geo-tagged data showed major activity frommetropolitan suburbs however activity was not restricted tothe regions geographically associated with each team. Wepresent a decay calculation methodology, using which wederive that activity died down on both Twitter and Face-book in a very similar manner. Such analysis can be usedto model events and study their penetration in social net-works. We analysed text for spot-fixing and found that userresponse to allegations about matches being fixed was cold.The complete analysis presented in this report, can be par-ticularly useful for studying events involving crisis or eventsof political importance having similar structure.
1. INTRODUCTION
Over the past few years, there has been an increase in theusage of Online Social Media (OSM) services as a mediumfor people to talk and discuss events. Genres of these eventsinclude politics, science, sports, entertainment, culture, phi-losophy and literature. Significant events like political elec-tions, riots, protests and crisis situations like predicted natu-ral disasters and terrorist attacks have been widely discussedon online social platforms. Our motivation has been to an-alyze an event having a defined schedule and pre-scheduledsub-events, which involves stakeholders and their support-ers, so that our analysis can be beneficial for such studies.Our analysis gives useful insights into the decay of events inthe digital world of social networks. ∗ Megha and Raghav did this work as a part of their summerwork at Precog (http://precog.iiitd.edu.in/). . Sports events trend on all online social networks. Onesuch prolonged and popular event was the Indian PremierLeague (IPL). It is a showcase for Twenty-20 (T20) cricket,which is the shortest form of play in the game. Top Indianand international players participate in IPL, making it theworld’s richest cricket tournament. IPL has been the sub-ject of several controversies involving allegations of cricketbetting, money laundering and spot fixing. The year 2013marked the sixth season of IPL. The league started on April3rd, 2013 and continued till May 26th, 2013. The nine teamswhich took part in the sixth season were Bangalore RoyalChallengers, Chennai Super Kings, Mumbai Indians, DelhiDaredevils, Kings XI Punjab, Kolkata Knight Riders, PuneWarriors India, Sunrisers Hyderabad and Rajasthan Roy-als. Every team name is associated with a city or state inIndia; like Rajasthan Royals is associated with the state ofRajasthan in the western part of the country, while DelhiDaredevils is associated with Delhi which is the nationalcapital. However, players for each team came from differ-ent parts of the country and cricket teams from all acrossthe world. All the teams competed against each other in 72matches which were followed by two qualifiers, one elimina-tor and the finals. During these 76 matches, there were dayswith no matches, a single match or two matches.IPL’13 particularly became a huge hit in Online SocialMedia [20]. Teams were ranked according to their presenceon popular online social networks. Chennai Super Kings ledwith an online social footprint of over 2.49 million followedby Kolkata Knight Riders with 2.14 million [17]. Particularplayers and their performances were also a major topic ofdiscussion throughout the league on OSNs. In this report,we analyze in detail the activity on Facebook and Twit-ter during the league. We attempt to study user behavior,amount of content, quality of content, user geography, teamsupport and spot-fixing of the league in detail. We correlatethe activity on Facebook and Twitter with the performanceof each of the teams during the season, its brand value andthe number of supporters. Analysis of the spot-fixing con-troversy gives an interesting insight into how a sub-eventwithin IPL aligns with the league as a whole.IPL shares structural similarity with a wide variety ofevents. Our decay analysis for IPL has been tested to workwell for events of all genres as shown in Section 4. Geo- https://2013.twitter.com/ a r X i v : . [ c s . S I] M a y nalysis of events, as we have done for IPL in this report withupcoming location based social networks like Foursquare and Facebook Places. Unlike many event based analysisthat have been done in the computer science community,our study is not restricted to Twitter. We have analyzedthan 2.5 million tweets and 700 thousand Facebook posts inthis study.
2. RELATED WORK
Popularity of content on the Web, like news articles [21],blog posts [13, 16] and posts in online discussion forums[2], vary on different temporal scales. For example, contenton micro-blogging platforms, like Twitter is very volatile,and pieces of content become popular and fade away in amatter of hours [9]. Blogging and micro-blogging networksshow temporal and topological patterns which largely ex-hibit power law behavior [14, 15]. Sitaram et al. have stud-ied the growth and decay of trending topics on Twitter [3].Temporal trends displayed by users have been analysed andthe decay factor has been found to decrease in a power lawfashion. Yang and Leskovec examined patterns of temporalbehavior for hashtags in Twitter [22]. They presented a sta-ble time series clustering algorithm and demonstrated thecommon temporal patterns that tweets containing hashtagsfollow.Many event-based studies have been done in the recentpast. Kairam et al. studied trending events during peak ac-tivity periods [10]. Identification, prediction, classificationand diffusion patterns of real world events have been widelyanalyzed (Kim et al. 2012 [12], Becker et al. 2010 [4], Ab-basi et al. 2011 [1]). Events of different genres have beencovered in these analysis. Zin et al. presented a knowledgebased event analysis framework for automatically analyzingkey events by using various social network sources in caseof disasters [23]. Sakaki et al. investigated real time inter-action of events like earthquakes on Twitter [19]. Guptaet al. studied crisis situations and their content credibilityfor events like the Mumbai blasts [8] and the Boston Blasts(2013).To the best of our knowledge, this is the first attempt tostudy in detail, an India-centric sports event on online socialmedia. Paridhi et al. studied the cross pollination of infor-mation on FIFA World Cup 2010 from Flickr, YouTube andFoursquare to Twitter [18]. Burnap et al. examined datarelated to the London Olympics 2012 to conceptualize therelationship between social actors, events and social mediafor understanding the dynamic reactions of populations [5].Clavio et al. have tried to identify the demographic char-acteristics of a sample of college football fans [6]. Kewal-ramani et al.have identified clusters in tweets for IPL 2011[11]. They labeled the clusters based on the teams that werea part of the league.
3. DATA COLLECTION
We used MultiOSN [7] for data collection. MultiOSNmonitors real world events on multiple online social media.The query terms have been listed in Table 6 (Appendix).This data had to be filtered because terms like ‘MI’ whichwere actually used to collect data corresponding to Mumbai https://foursquare.com/ Indians, a team in IPL, may be used in other contexts as wellsuch as referring to Michigan state or ‘mi’ in Spanish. Wefiltered the data using a more accurate and exhaustive listof hashtags related to IPL and keywords consisting of abbre-viations, names of teams, and players. We collected data ofover 4 million tweets from Twitter and over 3 million pub-lic posts from Facebook. Nearly 1.37% of the tweets and0.75% of the Facebook public posts were geo-tagged (Table1). Official statistics for IPL’13 claim that 6.7 million tweetswere recorded from one million users over a period of twomonths. Similar data for Facebook is not available. Ourfiltered dataset for Twitter comprises of nearly 2.6 milliontweets (around 39% of the claimed data) from about 485thousand unique users. Twitter FacebookTotal Tweets/Posts 4,895,784 3,753,499IPL Tweets/Posts (filtered) 2,627,197 774,186IPL Unique users 485,533 401,254Geo-tagged Tweets/Posts 36,126 5,827Start Date (mm/dd/yy) 04/01/13 04/01/13End Date (mm/dd/yy) 05/31/13 05/31/13Table 1: Descriptive statistics of the data collected.We also analysed the quality of data for Twitter. Figure1 represents the number of tweets per user for all the userswho tweeted about IPL. We can clearly observe a powerlaw curve for the number of tweets per user, which is inagreement with previous research work on Twitter [8, 14].Figure 1: Shows the distribution of number of tweets peruser on a logarithmic scale from the Twitter dataset.
4. DECAY METHODOLOGY
The way in which the activity of an event dies down is in-dicative of the impact an event has. The decay can be steepfor short-lived events while events which have an extendedpresence on OSNs depict a gentler fall. We analysed thedecay in activity on Twitter for the Texas fertiliser plantexplosion, Boston Marathon bombing, the Mumbai blastsand various other events. Figure 2 shows the fall in hourlytweet count and net decay factor for some of these events.We observed that the duration of decay may vary from afew hours to a few days. Another contributing factor can ben increase in the activity during the decay of an event. Wecan take into account all this information and calculate thenet decay factor.We obtained a graph for each event which depicted theactivity per hour on the vertical axis and time on the hor-izontal axis. From the plot, we identified the decay regioni.e. the region after the maximum activity. Starting fromthis peak value (marked with * in each plot in Figure 2), weconsidered all the data points except when the activity re-duced to less than 99% of the peak activity. This was definedas the threshold for insignificant decay. Beyond this limit,the activity is not included in the decay factor calculation.Within the decay, there can be some very steep peaks whichwe found could be linked to sub-events. In the event of anews update, there is a sudden surge in the activity but itdropped as quickly as it rose. These peaks do not correctlydepict the actual decay of the event and need not be consid-ered for the decay factor calculation. For our calculationswe removed the peaks before recursively dividing the decayregion into smaller regions. The divisions were made suchthat the R value for the fit of the line having an equationof the form of Equation (1) was more than 0.8. χ = α ln( t ) + β − (1)where χ represents activity, t represents time and α and β are constants.This particular equation has been chosen considering char-acteristics of plots for OSM activity and also by applyingvarious possible exponential and logarithmic models. Foreach region the value α/β is considered as its contributionto the net decay as it is directly proportional to the fall de-picted in the region. Similarly for regions having a positiveslope i.e. increase in activity, will have a negative α/β ratio.Calculations for the decay-growth pairs occurring simulta-neously are then aggregated to represent the pair as oneregion instead of two. The values obtained from each regionare then aggregated depending on the size of the region thustaking into consideration its contribution to ( α/β ) net .
5. ANALYSIS
The plots shown in Figures 3 represent data collected overtwo months during the IPL season. The peaks coincide withthe different matches and the last peak coincides with thefinal match. These plots clearly indicate lesser data for Face-book as compared to Twitter. This may be attributed to themore public nature of Twitter as compared to Facebook andalso to the absence of user data and non-public posts whichcannot be obtained using the Facebook search API. Wealso looked at day-wise plots and as shown in the plots fromthe Facebook dataset (Figure 4), these are clearly indicativeof the number of matches that were played on a particularday. The plots obtained from the Twitter dataset share thisproperty. We can infer from the plots that most activity hasbeen recorded during the matches and not after or before.A controversy arose during the season when three crick-eters were arrested on charges of spot-fixing (dishonestlydetermining the outcome of a specific part of a game be-fore it is played). Data for spot-fixing was filtered from theIPL Dateset starting 5:30 AM (IST) on the 14th of May till12:30 AM (IST) on the 8th of June and activity was mea-sured around specific keywords which have been listed inTable 7 (Appendix). 167,503 tweets (10.3% of the Twitter https://developers.facebook.com/docs/reference/api/ (a) Fertiliser plant explosion, Texas(b) Boston Marathon Bombing(c) Mumbai Blasts Figure 2: Decay in hourly tweet count along with net decayfactor for Texas fertiliser plant explosion, Boston marathonbombing and Mumbai blasts.IPL data) and 80,312 Facebook posts (6.3% of the FacebookIPL data) were obtained after filtering. These statistics indi-cate that spot-fixing was one of the major sub-events underIPL. Plots for activity corresponding to spot-fixing yieldedinteresting results.As can be seen in Figure 5 the frequency of peaks for thespot-fixing data is the same as the frequency of matches.We also found that in case of Twitter most peaks occurduring match timings. The significance of the plots can bebetter understood using the spot fixing timeline given in Ta-ble 2. The activity became significant soon after the threecricketers were arrested.We see that the peaks in the plots a) Tweets vs. Time (b) Facebook Posts vs. Time Figure 3: Hourly data collected for Twitter and Facebook over entire IPL. (a) Activity on days having one match (total 26 days).(b) Activity on days having two matches (total 25 days).
Figure 4: Daywise plots for Facebook data clearly depictinghigh frequency of posts during the match timings.coincide with important advances and updates in the spotfixing controversy. We also observed that the activity onboth platforms decays significantly from 26th of May, theday on which the final match was played, officially markingan end to this season. The spot-fixing controversy beganonly two weeks from the end of the IPL season, though itspercentage in the total data indicates that spot-fixing con-tributed to a significant fraction of the activity, 14th of Mayonwards. On the day of the final match (26th of May), spot-fixing accounted for 14.1% of the IPL activity on Twitter and 17.3% of the same on Facebook. This gives us good insightinto how sub-events work within events and how they canbe analysed together.May 16 Rajasthan Royals players Sreesanth,Ajit Chandila and Ankeet Chavanarrested for spot-fixing by Delhi PoliceMay 17 BCCI suspends former Royals player,Amit Singh, who was arrested as a bookieMay 18 Mumbai police link Ramesh, Ashok andKadam to the case; Seize Sreesanth andJiju Janardhan’s belongings from hotelrooms booked in their names at afive-star hotel in MumbaiMay 21 Vindoo Dara Singh arrested for allegedlinks with bookiesMay 23 Mumbai police team raids Meiyappan’sChennai residence, summons Meiyappan forquestioning, rejects his requestfor an extension till after IPL-VI finalsMay 24 Meiyappan arrives in Mumbai; He isarrested on charges of betting and cheatingMay 25 Mumbai police say Meiyappan passed oninformation about Chennai Super Kingsto bookiesMay 26 BCCI announces three-member commissionto investigate Meiyappan; ICC removesumpire Asad Rauf from the ChampionsTrophy squadTable 2: IPL spot-fixing timeline
We analysed the geo-tagged tweets and Facebook poststo better understand the geographic distribution of IPL ac-tivity on social media. Geo-tagged tweets only comprise asmall percentage of the dataset as seen in Table 3. Eventhough the results are in coherence with what might be ex- a) Twitter(b) Facebook
Figure 5: Plots from the spot-fixing dataset showing peaksin activity aligned with the spot-fixing timeline in Table 2.pected, the possibility of the geo-tagged dataset being bi-ased cannot be ignored. We plotted heatmaps to understandthe global distribution as well as the distribution within In-dia. Apart from India which accounted for 75.5% of thegeo-tagged tweets (27,271) and 89.8% of the Facebook posts(5,232) the list of other countries from where significant ac-tivity was seen included UK, UAE, USA, South Africa, Sin-gapore, Indonesia, Sri Lanka and Pakistan.Within India, the density is highest in the metropoli-tain areas. Delhi, Mumbai, Bangalore and Chennai are thecities with maximum activity as can be seen in Figure 6.Analysing addresses helped us further understand the dis-tribution between cities and upcountry. In case of twitterwe found that the top 50 most populated cities accordingto the census of India accounted for 69.95% of the tweets(19,075) and top 5 accounted for 46.98% (12,812) of thetweets. For Facebook the figures were 81.82% (4,281) and55.22% (2,889) respectively.Upcountry refers to the remaining 30.05% of the data i.e.tweets other than those from the 50 most populated cities inIndia. We went on to study the distribution for the upcoun-try data for Twitter and found that most states and regionswith high activity were among the more developed regions ofthe country and had high levels of literacy. The list of states http://censusindia.gov.in Twitter FacebookTotal Geo-Tagged Data 36,107 5,823Percentage in total IPL Data(%) 1.3 0.75Table 3: Geo-Tagged DatasetTeam Geo Tweets Local activityChennai Super Kings 7025 18.90%Mumbai Indians 5741 36.71%Royal Challengers 4730 16.74%Pune Warriors 1684 34.73%Sunrisers Hyderabad 2050 14.92%Delhi Daredevils 1869 18.56%Kolkata Knight Riders 2371 9.78%Kings XI Punjab 1449 4.07%Rajasthan Royals 2693 4.64%Table 4: Team-wise description of geo-tagged datasethaving contributed most significantly included Maharastra,Karnataka, Tamil Nadu, Delhi, Kerela and Gujarat. Therewere also states with no geo-tagged contributions on eitherof the social networks such as Jammu and Kashmir and Ma-nipur. Even in case of the upcountry data, a big chunk canbe attributed to suburbs of big cities, locations along impor-tant highways and to industrial belts. Along with internetpenetration, another contibuting factor to this skew is highprevalence of content in English, which is a hurdle for muchof rural India.We also analysed heatmaps of the geo-tagged data for eachteam separately and though most of the activity came fromthe home state and cities of each team, there was a spreadacross the metropolitans for all the teams. So people havebeen talking not only of their home teams but also of teamsfrom other parts of the country showing no regional or ge-ographical bias. We see in Table 4 that while in the caseof Mumbai Indians 36.71% of the geo-tagged tweets werefrom Maharastra, for Kings XI Punjab only 4.07% of thegeo-tagged tweets were from Punjab. It is worth notingthat both the teams from Maharastra, namely the MumbaiIndians and the Pune Warriors India have the highest con-tributions from their home states. Even though the activityfrom Chennai, Delhi and Banglore is huge they account forless than one-fifth of the geo-tagged tweets talking abouttheir home teams. In this section, we describe our analysis of linear depen-dencies between different sets of data using the Pearsonproduct-moment correlation coefficient. We studied corre-lations between team popularity on social media with theirbrand values and IPL rankings. Team popularity in this http://iplt20wiki.com/ipl-brand-value-up-by-4-to-3-03-billion-ipl-2013/4312/ Table 5 shows the corre-lations we found between different datasets. We can clearlyobserve high linear dependence among a team’s brand value,performance and its online presence. The correlation coef-ficient between the number of likes on a team’s Facebookpage and its brand value was calculated to be 0.92!Variables CoefficientFacebook Page Likes, Team Brand Value 0.92Facebook Page Likes, Tweets 0.83Tweets, Team Brand Value 0.79Tweets, Followers on Twitter 0.94Tweets, IPL Ranking 0.70IPL Website Tweet, Team Brand Value 0.73IPL Website Tweet, IPL Ranking 0.71Table 5: Pearson Correlations
Figure 6: Fall in activity on Facebook and Twitter depictingsimilar decay patterns on both networks.The plots shown in Figure 7 represent the fall in hourlytweet count after the finals of IPL between the teams Chen-nai Super Kings and Mumbai Indians. Prominent decay canbe observed within the first 20 hours for both the OSNs. Theplots clearly indicate a very similar decay pattern, which isalso evident from the decay factor we calculated for IPL fi-nals for both Facebook and Twitter. Applying the decaymethodology (Section 4) on the plot for Facebook resultedin a decay factor of 0.1942 while for Twitter it was 0.2022.The decay factors are comparable as expected. Individualdecay factors also indicate the steepness in fall of hourlytweets right after the finals. It is also interesting to notethat peak activity (the point in the plot from where decay We did a sentiment analysis on the content of tweets andFacebook posts using Linguistic Query and Word Count. We found that positive emotion in the text was higher thannegative emotion (Figure 8). Words related to anxiety hada surprisingly low count, merely 0.33% and 0.31% of theentire data for Twitter and Facebook respectively. Signif-icant percentages were found for social words as well aswords related to space and time. IPL’13 was the richestcricket tournament in the world and involved spot-fixingand money laundering incidents. Despite that, the percent-age for money words is very low for both the networks.Figure 7: Text analysis of IPL dataset for both Twitter andFacebook showing high percentage of ‘social’ and ‘positive-emotion’ words.We obtained the tweets and posts about spot-fixing andanalysed them separately using LIWC. The results obtainedfor Twitter are shown in Figure 9. The percentage for anxi-ety words was less for this data too. Percentages for positiveand negative emotions were less than the values obtained forthe entire IPL data. We could also observe a higher percent-age for money words being used in the context of matchesbeing fixed. To understand these results in a better man-ner and explain surprisingly low anxiety for the matchesdespite IPL’s significant presence in both offline and onlineworld, we analysed the tweet content manually. We foundthat IPL is more fun for people than anything else and theydidn’t care much about the matches being fixed. These re-sults show that IPL serves more as a means of entertainmentthan anything else. http://articles.timesofindia.indiatimes.com/2011-04-23/news/29466837 1 ipl-teams-nba-teams-indian-premier-leagueigure 8: Text analysis of IPL and spot-fixing dataset forTwitter. The graph shows less percentage of ‘positive-emotion’ and ‘anxiety’ words indicating that, despite hugefan-following and involvement, people view IPL as simpleentertainment.
6. SUMMARY
Online social media has become a platform to understandpublic opinion and sentiment. In this research work we anal-ysed the online activity during the Indian Premier League.We studied 2,627,197 tweets and 774,186 Facebook postsfrom 485,533 unique users on Twitter and 401,254 uniqueusers on Facebook. The plots for hourly tweets vs. timeclearly depicted the scheduled nature of the league. We anal-ysed the geo-tagged tweets to understand the geographic dis-tribution. Our results showed that 75.53% of the geo-taggedtweets and 89.85% of the geo-tagged posts were from Indiaand of those the most significant contribution came frommetropolitan suburbs and industrial belts. Decay analysisof the IPL activity yielded a net decay factor of 0.1942 forFacebook while for Twitter it was 0.2022. This clearly showsthat the activity on both social networks died down simi-larly. We found strong correlations between the brand valueof teams and how much they were talked of on Twitter (0.79)and Facebook (0.91). We also found that the correlations be-tween popularity of different teams on both the OSNs werestrong enough to say that they follow similar trends. Textanalysis of both the spot-fixing dataset and the completeIPL dataset using LIWC showed that even after the spot-fixing scandal, people viewed IPL as simple entertainmentand remained unaffected by the numerous allegations.To the best of our knowledge this is the first detailed anal-ysis of an India-centric sports event on online social me-dia. With more of such studies on the IPL and on othersuch events we can possibly find trends and draw parallelsto better understand user behaviour in case of significantsocial and political events. We could also further expandthis study using text analysis to find the number of sup-porters each team has and draw correlations similar to theones drawn in this report. We are currently collecting datafor IPL’14, and we plan to do a comparative analysis onceIPL’14 is over.
References [1] Abbasi M, Chai S, Liu H, and Sagoo K. Real-worldbehavior analysis through a social media lens. 2011. [2] Aperjis C, Huberman B.A, and Wu F. Harvesting col-lective intelligence: Temporal behavior in yahoo an-swers.
ArXiv e-prints , 2010.[3] Asur S, Huberman B.A, Szabo G, and Wang C. Trendsin social media : Persistence and decay. 2011.[4] Becker H, Naaman M, and Gravano L. Beyond trend-ing topics: Real-world event identification on twitter.
International AAAI Conference on Weblogs and SocialMedia (ICWSM) , 2011.[5] Burnap P, Housley W, Morgan J, Sloan L, Williams M,Avis N, Edwards A, Rana O, and Williams M. Socialmedia analysis, twitter and the london olympics 2012.2013.[6] Clavio G. Social media and the college football au-dience.
Journal of Issues in Intercollegiate Athletics ,2011.[7] Dewan P, Gupta M, Goyal K, and Kumaraguru P. Mul-tiosn: Realtime monitoring of real world events on mul-tiple online social media.
I-CARE Proceedings of the 5thIBM Collaborative Academia Research Exchange Work-shop , 2013.[8] Gupta A and Kumaraguru P. Twitter explodes withactivity in mumbai blasts!a lifeline or an unmonitoreddaemon in the lurking? 2011.[9] Java A, Song X, Finin T, and Tseng B. Why we twit-ter:understanding microblogging usage and communi-ties.
WebKDD workshop , pages 56–65, 2007.[10] Kairam S, Morris M, Teevan J, Liebling D, and DumaisS. Towards supporting search over trending events withsocial media.
Stanford University and Microsoft Re-search , 2013.[11] Kewalramani M. Community detection in twitter.
Uni-versity of Maryland , 2011.[12] Christen P Kim M, Xie L. Event diffusion patterns insocial media.
The Australian National University , 2012.[13] Kumar R, Novak J, Raghavan P, and Tomkins A. Onthe bursty evolution of blogspace.
World Wide WebConference , 2003.[14] Kwak H, Lee C, Park H, and Moon S. What is twitter,a socia network or a news media?
World Wide WebConference , 2010.[15] Leskovec J, Mcglohon M, Faloutsos C, Glance N, andHurst M. Cascading behavior in large blog graphs.
SIAM Conference on Data Mining (SDM), 2007 , 2007.[16] Mei Q, Liu C, Su H, and Zhai C. A probabilistic ap-proach to spatiotemporal theme pattern mining on we-blogs.
World Wide Web Conference , 2006.[17] Mishra O. Ipl on social media week 6: , 2011.[19] Sakaki T, Okazaki M, and Matsuo Y. Earthquakeshakes twitter users: real-time event detection by so-cial sensors.
WWW ’10 Proceedings of the 19th inter-national conference on World wide web , 2010.[20] Sharma R, Bhushan R, and Rathore V. Ipl strik-ing gold on social media platforms like twitter,google+ & youtube with corporate advertisers. 2013.http://articles.economictimes.indiatimes.com/2013-04-09/news/38404480.[21] Szabo G and Huberman B.A. Predicting the popularityof online content.
ArXiv e-prints , 2008.[22] Yang J and Leskovec J. Patterns of temporal variationin online media.
WSDM ’11 Proceedings of the fourthACM international conference on Web search and datamining , 2011[23] Zin T, Tin P, Hama H, and Toriu T. Knowledge basedsocial network applications to disaster event analysis.
International MultiConference of Engineers and Com-puter Scientists , 2013.
7. APPENDIX7. APPENDIX