Learning to link with wikipedia. Borau, K. Mi- Muff, S. Local modular- croblogging for language learning: Using twitter to train ity measure for network clusterizations.
Physical Review E communicative and cultural competence. In Proceedings 72 Newman, M. Modularity and community structure in networks. Short and tweet: experiments on recommending con- tent from information streams.
In Proc. Microblogging: A semantic and distributed approach. Clauset, A. Find- ing community structure in very large networks. Physical Ponzetto, S. Deriving a large scale Review E Collins, M. Unsupervised models for named entity classification. In Proceedings of the Joint Porter, M.
Cucerzan, S. Large-scale named entity disambigua- Sadikov, E. Correcting for missing data in information cas- CoNLL. Gomez-Rodriguez, M. Inferring networks of diffusion and influence. Krishnamurthy, B. A few chirps about twitter. In Proceedings of the first workshop on Sankaranarayanan, J. Twitterstand: news in tweets. Collective annotation of wikipedia entities in web text. Tweet the debates: understanding community annotation of uncollected sources.
What is workshop on Social media. In In Proc. Suh, G. Want to be retweeted? Information contagion: an retweet in twitter network. Weng, J. Twitterrank: finding topic-sensitive influential twitterers. In Proceedings Lerman, K.
Using a model of social of the third ACM international conference on Web search dynamics to predict popularity of news. In Proceedings of and data mining. White, S. A spectral clustering ap- Leskovec, J. In Proceedings of Hurst, M. Cascading behavior in large blog graphs. Yarowsky, D. Word-sense disambiguation using sta- Leskovec, J. Community structure in large networks: Natural clus- In Proceedings of the 14th conference on Computational lin- ter sizes and the absence of large well-defined clusters.
Related Papers. What is Twitter, a social network or a news media? By Jervy Velasquez. By Mark Meiss. By Rayne Man. Detecting and tracking political abuse in social media. By Bruno Goncalves. Share This Paper.
Background Citations. Methods Citations. Results Citations. Figures, Tables, and Topics from this paper. Citation Type. Has PDF. Publication Type. More Filters. Twitter and other microblogs have rapidly become a significant means by which people communicate with the world and each other in near realtime.
There has been a large number of studies surrounding … Expand. Who Will Retweet This? Detecting Strangers from Twitter to Retweet Information. In addition, behavioural intention was taken as part of predicting variables of actual retweeting behavior and was investigated with other motivation variables. Participants were actual Twitter users who were recruited on Twitter, and data collection was done through an online survey. The results showed that altruistic motivation among three prosocial motivations could predict actual retweeting behaviour through behavioural intention.
In addition, the differential effects of reciprocity motivation was found to vary depending on the sizes of the followers and the followees.
Already have an account? We focus here on a Twitter data set which we have processed using the approach above to generate user pro? We explore which of our four models best? We identi? After a short period of time, we had reached?
The snowball sampling has yielded a constrained set of users who make up on large connected component. It is not unreasonable to assume that we have a slice of Tweets that many of them are aware of, or at the very least is representative of a geographically focused set of Tweets they are likely to see.
Clearly this is a geographically biased sample, but it is also powerful because it is thusly constrained and is therefore quite useful as a deeper study in a geographic region. We have been monitoring these Twitterers since early September , and have tagged the?
The full tweet data set for this period consists of , Tweets. From here, we? This resulted in 11, users being selected and , Tweets being included.
We use this data to? For our retweet study we further down-selected to only consider users who retweeted at least? This down-selection resulted in us having users and retweets. Figure 4: General Model: y-axis is the ratio of retweets, and the x-axis is the number of minutes between a retweet and the original tweet.
As can be seen, this approximates a powerlaw distribution with a slope of? We next? We answer this question by going through the 79K retweets and computing for each of our models how likely that particular Tweet was to be retweeted. We identify the most likely model for each retweet by calculating the respective probabilities e. We sum up over all retweets how often each model was the most likely. We also compute, for each user, the overall best model for that particular user.
We do so by computing, for each model, the overall likelihood of seeing all retweets for a given user. We then compute, for each model, how many users were best explained by that model. Finally, we compute for each user a pro? This way we can get a behavior pro? While there may be statistical issues with? This distribution seem to follow a powerlaw distribution as we see in Figure 4 and when we? To do this, we computed, for each retweet, the similarity between the Tweet categories and the user pro?
Our similarity measure the cosine distance lies in the range [0 : 1], which we discretize into bins 0 through We then computed, for each bin, the ratio of retweets which fell into that bin separately for the pro? These ratios then make up the empirical distribution which is the?
To get a sense for whether these? This was done by computing the average similarity between any tweet and a user for the expected? Figure 5 shows the empirical distributions for the topicmodel and the pro? Note that the y-axis is a logscale. We show, for both distributions, the observed retweet? Although we can see that all distributions have a very high probability mass at 0 similarity, we see that the expected models had much higher masses at 0 and much lower probability mass as the similarity increased, whereas the observed models showed that there was a strong signal in the similarity.
We titled the paper that antihomophily wins because a 0 mass is still the general winner, although clearly similarity plays a signi? Model Number Ratio General 61 3. As we can see, the pro? Figure 5: Empirical Distributions of how well the topic of a retweet matches the pro? Note the log scale on the y-axis. See text for details. Model General Number of models used One Two Three Four 67 Table 4: How many user behaviors were best explained by a combination of one, two, three or all for models.
Table 1: How often was each model the most likely explanation for a retweet. The question we will answer below in the study is whether the content-based models in fact are an overall better? Table 1 shows, for each of the four models, how often it was the best explanation for an observed retweet. While the pro? It may be that some losses are insigni? It may also be that certain behaviors are more prominent with high-volume users, disproportionally making it show up.
To address this, we next explore whether this per-retweet behavior follows into general behaviors by users. As we can see, the qualitative behavior is roughly the same although the general model drops signi?
Most of the gain is in the topic-based model, bringing it almost to a tie with the recency-model. Finally, we wanted to explore if the above result shows a consistent behavior of users i. Table 3 shows how often models were used, on average, across all users. Clearly there is high variation within users from this average pro? For example, for all models it was the case that there was at least one user who never used that model as well as at least one user who exclusively?
All of these results indicate strongly that multiple models ought be used and explored when trying to understand why a particular information propagation or information diffusion pattern appears. We have shown that there is a signi? In particular, we studied a set of Twitter users over a period of a month and sought to explain the individual information diffusion behaviors, as represented by retweets, in this domain. We hypothesized that knowing more about the user and the content would allow us to develop richer models which would take pro?
We used these pro? We explored four retweeting models, two of which were based on user pro? We found that indeed the contentbased propagation models were better at explaining the majority retweet behaviors we saw in our data. This work is a?
0コメント