Predicting ReTweet Count Using Visual Cues

Transcription

1 Predicting ReTweet Count Using Visual Cues Huseyin Oktay * University of Massachusetts Amherst Amherst, MA, USA hoktay@cs.umass.edu Ethem F. Can University of Massachusetts Amherst Amherst, MA, USA efcan@cs.umass.edu R. Manmatha University of Massachusetts Amherst Amherst, MA, USA manmatha@cs.umass.edu ABSTRACT Social media allows rapid information diffusion, and serves a source of information to many of the users. Particularly, in Twitter information provided by tweets diffuses over the users through retweets. Hence, being able to predict the retweet count of a given tweet is important for understanding and controlling information diffusion on twitter. A tweet may contain just text or additionally a link to a web site or even an image. Since the length of a tweet is limited to 140 characters, extracting relevant features to predict the retweet count is a challenging task. However, visual features of images linked in tweets may provide predictive features. In this study, we focus on predicting expected the retweet count of a tweet by using visual cues of an image linked in that tweet in addition to content and structurebased features. The contributions of this paper are threefold. First, we introduce followers-to-friend ratio as another feature highly correlated with retweet count. Secondly, in addition to text-based features and structure-based features, we also extract visual features of an image linked in a tweet and we show that image features are highly correlated with the retweet count. Finally, we use such image features in predicting the retweet count of a given image, and empirically show that using visual cues increases the predictability of the retweet count of a tweet by 16%. Categories and Subject Descriptors H.2.8 [Database and Management]: Data Mining; J.4 [Computer Applications]: Social and Behavioral Sciences General Terms Algorithms, Experimentation Keywords Twitter, Retweet prediction, Visual Cues Equal contribution 1. INTRODUCTION Social media platforms enable users not only to share information with their peers but also to discover new information. Such information reaching to large populations has the potential to influence public opinion[2, 13], market share for new products[3] and the adoption of innovations[18]. To maximize the reach of a certain piece of information or product to large populations, researchers and marketers have become interested in using various different strategies on social media such as seeding information with influentials [12]. One way to maximize such efforts could be by predicting the diffusion of information on social media to quantify the expected exposure of different strategies. In particular, for example, the basic mechanism for the diffusion of such information over Twitter is through retweeting. Users can choose to retweet a particular tweet to share the content with their followers. Hence, understanding the mechanisms of information diffusion and predicting the retweet count beforehand are important tasks to quantify the expected reach of a particular piece of information. The prediction of the retweet count is a challenging task, partly because Twitter is a complex domain with many users interacting with each other by generating massive content on a daily basis. More than 200M tweets are generated per day by more than 100M users in different languages. Qualitative and quantitative studies have been performed to understand and predict retweetability. Boyd et al.[4], qualitatively reveal the factors affecting retweeting behavior as a conversational practice. In another study, Suh et al.[20] quantitatively analyze the factors that are highly correlated with the retweet count of a tweet. Their linear model suggests that having a hashtag, having a url, the number of followers and the number of friends as well as the age of the account is highly correlated with the retweet count of a tweet. One common constraint reflected in such studies is that Twitter is a limited domain where a tweet s content cannot be longer than 140 characters making it hard to come up with predictive content-based features. From the perspective of users, such a constraint is about the level of information that a tweet can contain. Hence, users use url links to websites, news articles, or pictures to at least refer to more content in their tweet. Users make use of url shorteners while they are tweeting so that they don t use too many characters (e.g., bit.ly), or they make use of photo hosting platforms combined with url shorteners to include images in their tweets. As users include more information

2 in their tweets by including pictures, we can make use of such additional information contained in the shared images in predictive models to estimate the retweet count for a given tweet. In this paper, we focus on tweets that contain links to images shared through twitpic.com, and extract correlated low-level and high-level image features in those tweets. We, then, use these correlated visual cues to increase the accuracy of our predictive model for retweet count. Our low level features are global color histograms and GIST features [11, 16] while our high level features use the set of responses to specific object detectors. We, then, compare three different regression techniques (linear regression, support vector machines and random forests) over these features as predictive models to estimate the retweet count of a given tweet. Our contributions in this paper are three-fold. First, we introduce followers-to-friends ratio as another feature highlycorrelated with the retweet count, and show that random forest regression model including this new feature along with additional content- and structure-based features suggested in the literature gives a statistically significant improvement in performance. This new combination of features serves as our baseline set of features in our experiments. Secondly, we augment our baseline features with image features either color, GIST or high-level features and show that the Pearson s correlation of some of the visual features is highlycorrelated with the retweet count. Finally, we show that using such image features in our regression models we can substantially increase the accuracy of predicting the retweet count for a given tweet. The rest of the paper is structured as follows. In section 2, we discuss the existing related work both about predictive modeling in Twitter as well as extracting image-based features from images. In section 3, we introduce the data set we used, and explain content-, structure and image-based features used in our analysis. In section 4, we formulate the problem, and explain the experimental design. That is followed by the presentation of experimental evaluation of different regression models with additional image-based features. In section 5 and 6, we discuss the strengths, weaknesses and possible improvements of our results and then we conclude the paper. 2. RELATED WORK A number of qualitative and quantitative analysis have been performed towards understanding information diffusion and influence on Twitter. Kwak et al.[14] and Cha et al.[7] qualitatively compared different definitions of influence and mainly showed that influence ranking depends on the definition. Boyd et al.[4] formulated retweeting as a conversation practice to qualitatively understand the underlying mechanism for retweeting. Yang et al.[23] compared diffusion structures on Twitter to diffusion structures on web logs to identify qualitative differences and similarities. Quantitatively, Suh et al.[20] used generalized linear models and identified content-based and structure-based features that are significantly correlated with retweet count such as having a link in a tweet and the number of followers of the user who posted the tweet. Yang et al.[24] used survival analysis to predict the scale, speed and range of tweets by focusing on the diffusion of certain topics on Twitter network. They also used content- and structure-based features in their predictive model. Bakshy et al.[2] focused on predicting the influence of individuals on Twitter by considering the cascades of shared urls. In their predictive model they use manually labelled content-based features along with structural features and the past influence of a user to predict influence from a user perspective. Other problem formulations from a user s perspective include the following papers which predict the retweet count of a user. Zaman et al.[26] used a collaborative filtering model utilizing content- and structure-based features. Yang et al.[25] used a factor graph model. Peng et al.[17] used conditional random fields with temporal features on top of context- and structure-based features. Uysal and Croft[22], also predicted the likelihood of a retweet by a user on a given tweet with a logistic regression model and used the information to rank the tweets according to their interest to a user. In this paper, we quantitatively predict the retweet count from a tweets perspective. In other words, for a given tweet with a set of content- and structure-based features, we estimate the retweet count of a tweet. Additionally in our models, we use image features (both low-level and high-level) features of images in tweets along with the suggested contextand structural-based features suggested in the literature as being highly correlated with retweet count. Extracting visual features for images has been well studied. Low level features that are traditionally used include color histograms and GIST features for image retrieval [11, 16] Recently, high-level representations have been shown to work better than low-level representations due to their capability of providing semantically richer representations[15]. In high-level representations the major focus is on objects, scenes, and attributes rather than the distribution of colors or textures in images. High level representations are created using low level representations. For example Classemes is one such representation in which an image is represented as a histogram of concepts where the concepts are created by running SVM classifiers on images [21]. ObjectBank [10, 15] is another high-level representation where each image as represented as a histogram of object detector responses. Specifically, a number of object detectors are trained and responses to such object detectors are detected at different detection scales (12) and spatial pyramid levels. In a spatial pyramid an image is subdivided hierarchically. In this particular case the image is subdivided into four and into sixteen regions so in total = 21 regions/grids are computed. Given that there are 177 object detectors the histogram is of dimension The idea behind ObjectBank and classemes is that while individual object detectors may not be very accurate (as low as 20% accuracy for some), a representation using a histogram of objects or concepts can be shown to be useful for many classification tasks. We experiment with both low level features (global color histograms and GIST) and an ObjectBank type representation by doing max pooling over all detection scales and spatial pyramid levels to obtain one response per object detector per image for a histogram of dimension 177.

3 Table 1: Summary of features Name Value Content-Based hashashtab binary retweetcount Structure-Based followerscount friendscount followersfriendsratio age statuscount favoritescount listedcount Image-Based colorhistogram cf 0,..., cf 511 GIST gf 0,..., gf 511 objectdetectors obf 0,..obf DATA 3.1 Collection We crawled Twitter data using the Twitter sampling API [1] between March 29 th and April 10 th, The Twitter sampling API samples roughly 1% of all tweets from the Twitter firehose (i.e., full feed of public tweets). A random sample preserves the distribution of tweets. Among almost 33 M tweets, we focus on tweets that include links to pictures. We sampled the tweets that shared a picture using twitpic.com which is the most commonly used picture sharing platform on Twitter[20]. Among 140K tweets in our sample with a link to pictures, we were able to extract visual features for more than 100K of those tweets (some of the links did not contain images and some of the images were GIF files, and hence we eliminated those tweets from our sample). We first gather content- and structure-based features as a baseline as suggested in various different studies in the literature[20, 24, 26, 4, 17]. Then, we augment the features with additional visual features from images linked in tweets. 3.2 Features In our predictive models we focus on three types of features; content-based, structure-based, and image-based features. We construct our baseline experiments focusing on contentand structure-based features(i.e., baseline features). Our approach considers image-based features as well as the baseline features. In Table 1 we summarize the features and in the next sections we give more detailed information Content-Based Features hashashtag is a binary feature that represents the presence or absence of a hashtag in a tweet. Our target variable of interest is the retweetcount of a given tweet. As shown in Figure 1(a) and also suggested in the literature [23], the retweetcount has a power law distribution. To avoid the bias from one data instance with a high number of retweetcount, we transformed the data to log scale, and use the log value in our predictive models. To show that the distribution is a power low, we fit a line to log values of retweet count and we find that the slope of the fitted line is -0.35, and the r- squared value is 0.78 (r-squared can be interpreted to mean that 78% of the variance in the data is explained by the fitted line) Structure-Based Features followerscount is the total number of followers of the owner of a tweet. When a user sends a tweet, all the followers of that particular tweet get that particular tweet in their news feed. Hence the more the followers a user has, the wider the exposure of tweets of that user. friendscount is the number of friends that the owner of a tweet (i.e. the number of people that the owner follows). A user gets all the tweets from her friends. Hence, the more friends a user has, the more tweets a user gets. To again avoid biasing the scores towards users with a high number of followers in our models, we also transform the values of these features by using the log of the feature values. In Figure 1(d) and 1(e), we show the log transformation and the corresponding fit of a line in the log-scale (respectively slope:-0.13, R-square:0.45 and slope:-0.29, R-square:0.61). One might argue that followerscount and friendscount do not follow a perfect power law distribution, however, we think we can partially correct for the skew in the data by transforming it into log scale. followersfriendsratio is a ratio of the number of followers to the number of friends for the owner of a tweet. Such a ratio can be considered as a measure of the activity of a particular user. A user with a high value of the ratio (i.e., a high number of followers and a low number of friends) might be a user who uses Twitter to get her message out but not necessarily to get information from her friends. Conversely, a user with a low value of the ratio (i.e., a low number of followers and a high number of friends) might be a user who uses Twitter to get information from her friends but not necessarily to get her message out. A user with a roughly equal number of followers and friends might be a user who considers Twitter both to get her message out and to get new information. age is the total number of days the user has been active until a tweet is posted. Another structure-based feature is status- Count which is the total number of tweets of a user until the particular tweet is created (i.e., the total number of tweets tweeted by a user). As shown in Figure 1(f), status count is also a power law distribution (slope:-0.24, R-squared:0.70). Hence, we transform the feature values into log scale. favoritescount and listedcount are relatively recent features that have become available on Twitter data. favoritescount is the number of tweets a user marked as favorite and as shown in Figure 1(b), it is also a power law distribution (slope:-0.39, R-squared:0.73). listedcount is the number of public lists that a user is a member of. As shown in Figure 1(c), it also follows a power law distribution (slope: -0.50, R-squared:0.77). Hence, we transform both of the feature values into log scale Image-Based Features We also consider two different global low-level features. First, color histograms provide information about the distribution of color intensities. Second, GIST descriptors provide a set of perceptual dimensions such as naturalness which represents the major spatial structure of an image. In order to see the effect of low-level features we compare the performance of our predictive methods with color histograms and GIST descriptors as our image features separately. We ex-

4 tract color histograms as our image features in the following way. For each channel of a color image, i.e. red, green, and blue,, we uniformly quantize the range of color intensities into 8 distinct values. Therefore, we have a histogram of 512 bins in total (i.e., 8 bins per channel), and each bin represents a dimension in our feature vector. We compute a GIST descriptor for each image in the following way. The image is subdivided into 16 regions (i.e., 4 by 4 grids) and there are 8 orientations at each scale. The number of scales is set to 4 in our case. Therefore, in total there are 512 features(i.e., 16 regions x 8 orientations x 4 scales) that form a GIST descriptor for an image. We use an existing tool 1. The set of responses of individual object detectors {obf 0,..., obf 176} are used in our predictive models as well. In order to extract the responses of a number of object detectors we use the publicly available software provided by [15] 2. The standard output of the provided code are the responses of a number of object detectors at different detection scale and spatial pyramid levels. Here, we are interested in detection rather than classification so that we can correlate object detection responses with the retweet count. For each object detector we select the maximum value among the responses at different detection scales and spatial pyramid levels. In this way, we only have one, which is the maximum for that particular object, response for each object detector. In other words, we use max pooling over the responses at different scales and levels for an image. In our object bank, there are object detectors for 177 different objects such as dog, car, and table. Note that for each object we have a number output by the corresponding detector. Having computed and max-pooled the responses, for each image we have a 177-dimensional histogram. 4. EXPERIMENTS 4.1 Problem Formulation We formulate the problem of predicting the retweet count of a given tweet as a regression problem. Given a set of features F, and a target variable T, we experiment with different regression methods to predict a target variable. Our target variable is the logarithm of the retweet count for a given tweet. Our feature set F contains all the features explained in section 2. In our experiments, we focus on three different types of regression; linear, SVM with a Gaussian kernel, and random forest [5, 8]. Linear regression is one of the simplest and basic regression techniques. SVM regression is based on the same idea as SVM classification. We employ a nonlinear kernel - a Gaussian kernel. The last technique we focus on is random forest regression. It consists of a number of decision trees. Randomly subsampled features are used to train a decision tree and the rest of the features are then used to estimate the error to provide a better tree. As we will see later the choice of regression technique can give very different results. 4.2 Experimental Design In our experiments, we compare two different cases: baseline and the baseline augmented with image-features. As a baseline we use content- and structure-based features as explained in section 3.2. Such features include almost all the features reported in the literature as being highly correlated with the retweet count. Different from the literature, we also include followerstofriendratio in our baseline feature set which is highly correlated with the retweet count as shown in Figure 2(b). We augment our baseline features with image features either color, GIST or high-level features as as described in section We show the Pearson s correlation of image features we extracted with retweet count in our dataset. In Figure 2(a), we show the correlation values for the baseline features. In Figure 2(c), we show the correlation values for color features. Interestingly we observe that c511 which corresponds to white colors in the image has a high correlation with retweet count. In Figure 2(d), we show the correlation values for GIST features. And finally in Figure 2(b) we show the correlation values for object-based features with the retweet count. Especially for object-based and color correlation values we observe that some of the image-based features are highly correlated with the retweet count, which might suggested that addition of such features to a predictive model should increase the accuracy. Among 100K tweets, we create a training set by randomly sampling 10K tweets and we sample a different set of 10K tweets to create the test set. We train our model on the training set, and we evaluate the performance of our model on the test set. We replicate this sampling process 100 times and we evaluate our results using the root square mean error (RMSE) metric. For this metric, the lower the value, the better the performance. All the results in this paper are based on this evaluation metric. In our experiments, we make use of the R statistical software package. For SVM we use kernlab version package that is already implemented in R. For random forest regression we use the randomforest package version which is also implemented in R. We use default parameter settings. 4.3 Experimental Results In Figure 3, we compare the baseline and our approach for different regression methods (i.e., linear regression, SVM, random forest respectively). At a high-level, independent of the regression method, visual features extracted from object responses improve the prediction accuracy of the retweet count for a given tweet. Low level features also improve the prediction accuracy for some classifiers. We observe the most improvement with the random forest. Experiments with baseline features provide a root mean square error (RMSE) score of 1.743, 1.722, and in log scale and 5.713, 5.599, and in linear scale) for linear, SVM, and random forest regressions respectively (note that a lower RMSE is better). With our approach using object based features the RMSE scores down to 1.703, 1.559, and (5.489, 4.753, and in linear scale). The results with our approach are statistically significantly different from the results of the baseline with three different regression techniques 2. To reduce the variance, we slightly modify our approach to discard features that are not correlated enough with retweet

5 (a) retweetcount Distribution (b) favoritescount Distribution (c) listedcount Distribution (d) followerscount Distribution (e) friendscount Distribution (f) statuscount Distribution Figure 1: Distribution plots. count (i.e., below a threshold value=0.05). We labeled this case as cutoff in Figure 3. For object features, from a total of 186 features, by thresholding we use only 60 features that are correlated enough with retweet count. Our conclusion is that we observe slight improvements by thresholding features with linear regression, whereas the performance is not statistically significantly different for SVM and random forest models. Figure 3: Prediction results of the models with the baseline features compared with the models with baseline features and object-based features (also cutoff experiments) on three different regression techniques. To understand which features help we compute the correlations of the baseline features and the ObjectBank type features. with the retweet count and sort the values in terms of significance. In Figure 2(b), we provide the top 20 most correlated features with the retweet count. The most correlated feature is followerscount. This outcome is not surprising since the more followers, the more the users a tweet reaches, and hence the higher the retweet probability. Surprisingly there are 16 image-based features in the list of top 20 most correlated features and many of the structure-based and content-based features don t appear in the list of top 20. This figure shows why visual cues improve the accuracy of prediction models. The only feature which is negatively correlated is the age feature. In Figure 4, we compare the results of baseline with the lowlevel and high-level image image-based features with random forest regression. The results show that using a random forest, low-level image-based features also improve the prediction accuracy of retweet counts. The prediction model using low-level color histogram features perform as well as high-

6 (a) Correlation of retweet count with the baseline features. (b) Correlation of retweet count with the baseline features + object-based image features. (c) Correlation of retweet count with the baseline features + color-based image features. (d) Correlation of retweet count with the baseline features + GIST features. Figure 2: Top-20 Correlations for different sets of features; baseline, color, GIST, and object (followers: followerscount, ratio: followersfriendsratio, favorites: favoritescount, obf14: object-based feature-14, cf82: color-based feature-82, and gf94: GIST feature-94).

7 Table 2: T-test scores of the results of the models with the baseline and our approach with three different regression techniques (lb: lower bound, ub: upper bound). linear SVM random forest lb ub p-value lb ub p-value lb ub p-value Baseline vs Object < < <0.01 Baseline vs Cutoff < < < 0.01 Object vs Cutoff < Table 3: T-test scores of the results of the models with baseline, low-level, and high-level features with random forest regression (ub: upper-bound, lb: lower-bound). lb ub p-value Baseline vs GIST <0.01 Baseline vs Object <0.01 Baseline vs Color < 0.01 Baseline vs Cutoff <0.01 GIST vs Object <0.01 GIST vs Color <0.01 GIST vs Cutoff <0.01 Object vs Color Object vs Cutoff Color vs Cutoff Figure 4: Prediction result of the models with the baseline features compared with the models with the baseline features and different image-based features when random forest regression is used. level object features. The prediction model with low-level GIST features outperforms the accuracy of the model using just baseline feature. However, GIST features are not as good as color histograms or object-based features. Note that all image-based features are combined with the baseline features. The t-test scores are provided in Table DISCUSSION When we consider the results of random forest regression, the best models include baseline features combined with color-based features or baseline features combined with objectbased features. The model which combines baseline features with GIST descriptors is slightly worse and the model with baseline features is substantially worse and is the worst performer. Here, we claim that visual cues independent of whether they are low-level or high-level features improve the predictive power of our model using random forests. However, when we consider different regression techniques the results are not the same. The models with color-based features + baseline or GIST descriptors + baseline are not better than the model with baseline features alone when using linear regression or SVM. Models with object-based features always performs better than the model with baseline features independent of the regression method we use. Objectbased features provide more robust results than other lowlevel image features independent of the regression technique. With random forest regression, we get an RMSE value of in log scale, by improving the accuracy over baseline 16%. Since the difference in log-scale, is division in linear scale, our RMSE value for random forest corresponds to 3.66 times, and the baseline corresponds to 5.49 times in linear scale. For example, if the actual retweet count of a tweet is 100, baseline predicts a value roughly between 18 and 549, whereas random forest model predicts between 27 and 366. Although such prediction results may not seem perfect enough, the results significantly improve by including visual features from images and that is the main purpose of this paper. We conjecture that the more the accurate features extracted from images, the more accurate the performance of prediction of the retweet count wiil be. Currently, we consider responses to object detectors as one of our image-based features, and we observe that they are more correlated with retweet count than some of the structure based features. Responses of object detectors such as flower and ferris-wheel are positively correlated with retweet count.on the other hand, responses of object features such as building and blind are negatively correlated with retweet count. Besides, the responses for desk and car detectors are not significantly correlated with retweet count. Note that the responses that are used in object-based features are not binary values indicating the presence or absence of an object. However, they measure the possibility of an object being detected in an image at any detection scale and any spatial pyramid level. Even though, ObjectBank is one of the state-of-the-art methods, the performance of the detectors might not be as accurate as desired. Besides, using only the responses of object detectors would not be sufficient for a strong detection model. Word-phrases, and attributes might be used to increase the performance of detection as proposed in [9, 19] and thus might also improve the performance of our predictive model. When we consider the color-based

8 features, we only focus on simple color intensity distributions over the images. However, more complicated methods as suggested in [6] can be used to increase the correlation between images features and retweet count. We also find that followersfriendsratio, is highly correlated with the retweet count (see Figure 2(b)),and improves the baseline performance significantly when used with random forest model. This feature as a measure of user behavior in twitter can be used in many other predictive modelling tasks. Furthermore, for the tweets that contains a link to a web-site, the information in such sites could also be exploited to improve the accuracy of predicting retweet count. 6. CONCLUSION In this study we focus on the task of predicting the retweet count of tweets. We focus on content- and structure-based features as well as image-based features for our predictive models. The baseline features, with content- and structurebased features, provides an RMSE of 1.743, 1.722, and in log scale and 5.713, 5.599, and in linear scale by using linear, SVM, and random forest regression respectively. Our approach in which we augment the set of baseline features with image-based features improves the prediction accuracy of the models and provides RMSE of 1.703, 1.559, and in log scale and 5.489, 4.753, and in linear scale by using linear, SVM and random forest regression respectively. Among different sets of visual cues, we find that predictive model with object-based features provide more accurate results and is more robust than other image-based features to a change in the regression technique used. The addition of image-based features increases the accuracy of predicting which tweets are most likely to be retweeted. We conjecture that the more accurate information exploited from the links in tweets, the more accurate the prediction models should perform. 7. ACKNOWLEDGMENTS This work was supported in part by the Center for Intelligent Information Retrieval and in part by NSF grant #IIS Any opinions, findings and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect those of the sponsor. We would like to thank Ibrahim Uysal for his support. 8. REFERENCES [1] Twitter developers. March [2] E. Bakshy, J. M. Hofman, W. A. Mason, and D. J. Watts. Everyone s an influencer: quantifying influence on twitter. In Proceedings of the fourth ACM international conference on Web search and data mining, WSDM 11, pages 65 74, New York, NY, USA, ACM. [3] F. M. Bass. A new product growth for model consumer durables. Manage. Sci., 50(12 Supplement): , Dec [4] d. boyd, S. Golder, and G. Lotan. Tweet tweet retweet: Conversational aspects of retweeting on twitter. In Proceedings of HICSS-43, pages 1 10, Kauai, HI, January IEEE Computer Society. [5] L. Breiman. Random forests. Machine Learning, 45(1):5 32, [6] G. J. Burghouts and J. M. Geusebroek. Performance evaluation of local colour invariants. Computer Vision Image Understanding, 113:48 62, [7] M. Cha, H. Haddadi, F. Benevenuto, and K. P. Gummadi. Measuring User Influence in Twitter: The Million Follower Fallacy. In Fourth International AAAI Conference on Weblogs and Social Media, May [8] R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification (2nd Edition). Wiley-Interscience, 2 edition, [9] A. Farhadi, I. Endres, D. Hoiem, and D. Forsyth. Describing objects by their attributes. In Proceedings of the CVPR, pages , Los Alamitos, CA, USA, [10] L. Fei-Fei and L. J. Li. What, where and who? telling the story of an image by activity classification, scene recognition and object categorization. In Computer Vision: Detection, Recognition and Reconstruction, pages , [11] M. Flickner, H. Sawhney, W. Niblack, J. Ashley, Q. Huang, B. Dom, M. Gorkani, J. Hafner, D. Lee, D. Petkovic, D. Steele, and P. Yanker. Query by image and video content: the QBIC system. IEEE Computer, 28(9):23 32, [12] M. Gladwell. The Tipping Point: How Little Things Can Make a Big Difference. Back Bay Books, Jan [13] E. Katz and F. Paul. Personal Influence: The Part Played by People in the Flow of Mass Communications. Transaction Publishers, [14] H. Kwak, C. Lee, H. Park, and S. Moon. What is Twitter, a social network or a news media? In WWW 10: Proceedings of the 19th international conference on World wide web, pages , New York, NY, USA, ACM. [15] L. J. Li, H. Su, E. P. Zing, and L. Fei-Fei. Object bank: A high-level image representation for scene classification and semantic feature sparsification. In Proceedings of the NIPS, Vancouver, Canada, [16] A. Oliva and A. Torralba. Modeling the shape of the scene: a holistic representation of the spatial envelope. Internation Journal of Computer Vision, 42(3): , [17] H.-K. Peng, J. Zhu, D. Piao, R. Yan, and Y. Zhang. Retweet modeling using conditional random fields. Data Mining Workshops, International Conference on, 0: , [18] E. M. Rogers. Diffusion of Innovations, 5th Edition. Free Press, 5th edition, Aug [19] M. A. Sadeghi and A. Farhadi. Recognition using visual phrases. In Proceedings of the CVPR, pages , Colarado Springs, USA, [20] B. Suh, L. Hong, P. Pirolli, and E. H. Chi. Want to be Retweeted? Large Scale Analytics on Factors Impacting Retweet in Twitter Network. In Proceedings of the 2010 IEEE Second International Conference on

9 Social Computing, pages , Aug [21] L. Torresani, M. Szummer, and A. Fitzgibbon. Efficient object category recognition using classemes. In Proceedings of the ECCV, pages , Greece, [22] I. Uysal and W. B. Croft. User oriented tweet ranking: a filtering approach to microblogs. In CIKM, pages , [23] J. Yang and S. Counts. Comparing information diffusion structure in weblogs and microblogs. In ICWSM [24], pages [24] J. Yang and S. Counts. Predicting the speed, scale, and range of information diffusion in Twitter. In 4th International AAAI Conference on Weblogs and Social Media (ICWSM), pages , May [25] Z. Yang, J. Guo, K. Cai, J. Tang, J. Li, L. Zhang, and Z. Su. Understanding retweeting behaviors in social networks. In CIKM, pages , [26] T. R. Zaman, R. Herbrich, J. Van Gael, and D. Stern. Predicting Information Spreading in Twitter. Computer and Information Science, pages 1 4, 2010.