tions can be realized. (Gimple et al., 2011) for exam-ple, alreadydevelopedaPOS-taggerespeciallyfortheneeds of Twitter. Summarizing the results of all cate-gorized user Tweets, leads to a percentaged classiﬁca-tion of a user. There are several techniques availableand approved for realizing this classiﬁcation task. Re-ferred to section 3.1 SVMs can be used for such a task as applied by (Nakagawa et al., 2001). But there areseveral other ways for accomplishing this classiﬁca-tion behavior like using Bayesian approaches (Gold-water and Grifﬁths, 2005). Future testing and evalua-tion will give clarity about the best way for realizingcategorization of Twitter users. One possibility forevaluation is presented by (Chen et al., 2009).
3.3 Additional Ratios
In addition to measuring the similarity of
attributes, regarding the afﬁliation of a userinto a category, several other ratios for determiningthe importance of a users recommendation are usedto sharpen the prediction accuracy. The following ra-tios are legit candidates for additionally inﬂuencingwether a user within a topic related bubble will berecommended or not.
is the amount of Tweets a Twit-ter user is ﬁring within a deﬁned period of time.
. The more followers a userhas, the more inﬂuence or credibility one mightposses. On the other side, if a user has very fewfollowers, but is following a huge amount of otherusers, might hint to a
The amount of retweets a users Tweets have, indi-cates the amplitude a users reputation has.
If an observed user isn’t connected with the innercircle bidirectionally, this denotes a non friend-ship but a sheer interest related relationship.
Clients will have the possibility to rate recom-mended users or Tweets as ”interesting” or ”notinteresting” for a speciﬁed category. By com-paring users, which are rated as interesting withpotential recommendations for a
,similarity between those, can also inﬂuence theusers overall rating score within a bubble.These ratios could help to sharpen the selectionof recommended Tweets and Twitter users. However,the main task regarding applying these ratios, is toﬁnd an appropriate weighting scheme for every ratio.
http://www.makeuseof.com/dir/blastfollow-mass-follow-twitter-users/ (April 2012)
Recommendation decisions are made by calculatingratings for each potentially interesting user, based ontheir category classiﬁcation and the additional ratios,mentioned in section 3.2 and section 3.3. Subse-quently, category classiﬁcation of an active serviceuser, is compared to the classiﬁed categories of po-tentially interesting other users. In advance, all addi-tional ratios have different weights, which will ﬁnallyinﬂuence the position of a user in the ﬁnal recommen-dation list. Deﬁnite values for those ratios have to befound during development and test runs of the systemand therefore, can’t be predicted previously.
4 DEMO APPLICATION
Thought Bubble Server
will be implementedin Python and runs on an
web server. Figure2 visualizes the potential infrastructure of this system.
ClassiﬁcationWorkerThreadsTweetCollectorClients (iOS,Web, etc)RaterExternalInternalDatabaseOperationsThreadDatabaseWrapper
Figure 2: Thought Bubble infrastructure.
Twitter related API calls, which affect or are sig-niﬁcant for the classiﬁcation and recommendationtask, are processed and cached by the
server. The REST API acts as junction between theTwitter REST API and the client. All requests whicharen’t affecting the functionality of the Thought Bub-ble system, are directly processed by the TwitterREST API. When the system has completed catego-rizing and rating of potential recommendationsfor theﬁrst time a user starts to use this service, the systemstarts to enrich the Twitter stream with Tweets fromrecommended persons. Recommendation of singleTweets is based on the inﬂuence a Tweet has had dur-ing classiﬁcation of a certain user. Thought Bubbleclients can be used just like usual Twitter clients forreading ones personal Twitter stream, tweeting or di-rect messaging. However, the big difference is that