Analyzing Image Popularity
on a Social Media Platform
Esther Fiolet
Student no. 10574050 Final version July 25, 2014Supervisor: Stevan Rudinac Signature:
Second Examiner: Frank Nack Signature:
Master Thesis Information Science Human Centered Multimedia University of Amsterdam Faculty of Science
Analyzing Image Popularity
on a Social Media Platform
Esther Fiolet
HCM, Information Science University of Amsterdam Amsterdam, the Netherlands
estherfiolet@live.nl
ABSTRACT
Social media is a rapidly growing segment of the Internet. A large amount of user-generated content is uploaded daily, but its popularity differs greatly. Understanding what makes certain content popular can be of great interest to e.g. discover patterns in collective user behavior or for marketing purposes. This study attempts to reveal a range of factors that may influence image popularity on social media, focusing on photo-sharing platform Instagram. The results are obtained by performing an analysis of a set of images, by conducting user studies about behavior of Instagram users and by investigating the importance of image content. Results indicate the social context of an image plays a significant role in image popularity on Instagram.
General Terms
Human Factors
Keywords
Social Media, User Studies, Image Popularity, Instagram
1. INTRODUCTION
In the past decade, social media has become increasingly important to create, share and rate content. Some popular examples are Facebook, YouTube, Twitter, Instagram and Pinterest. These platforms cover a large segment of the Internet. YouTube for instance, reports that 100 hours of video is uploaded to their website every minute [23] and Instagram reports an average of 55 million images uploaded to their service daily [24]. Social media platforms can be construed as a form of collective wisdom [1] because they provide a rich repository of peoples’ opinions on a large number of topics. Users are stimulated to participate in sharing their thoughts by ‘like’ & ‘dislike’ tools, up-votes and tagging and commenting functions. This makes user-generated content typically enriched with metadata like annotations, geo-tags, comments and ratings.
Due to their popularity, social platforms get increasingly exploited for marketing of content: movie studios place trailers on YouTube, U.S. presidential candidates run political campaigns online and artists promote their work on networks and blogs, hoping to reach an audience of millions of people [2]. However, popularity of content differs greatly. To describe this problem we can refer to the long tail phenomenon [3], where a relatively small number of content reaches a broad audience and receives lots of attention, while the majority of content (the tail) receives little attention or none. Understanding why certain content becomes more popular can be of great interest. It can discover properties and rules about collective user behavior and it can help content creators and online marketers to better understand popularity to drive their future actions [4]. However, user behavior and their attention span are known to be unpredictable and therefore it can be hard to discover patterns [5]. When it comes to popularity of
topics, a study [6] suggests there is considerable evidence that popularity partly depends on novelty and competition of new topics. When new content gets propagated through the network, it takes places of earlier topics of interest causing limited attention of users, which makes the content to soon become invisible. Furthermore, popular and viral content can benefit from the rich get richer dynamic, which means the more popular it becomes, the more views it is likely to attract [7]. Recent studies, such as [8], suggest a high degree of correlation between popularity, importance or interestingness of content and its number of views, ratings and comments.
A number of studies [9] [8] [15] have focused on revealing evolutionary patterns of user-generated content and creating models to predict popularity. This study aims to determine a range of factors of social media content that may influence popularity. As a test-bed for this study, Instagram is used. This is a rapidly growing photo-sharing application used on mobile devices. It can serve as an interesting source of data because it contains large amounts of user-generated images that are peer-rated on large scale.
The main contributions of this study are the wide range of factors that are investigated and the user studies approach, which could contribute to new insights concerning content popularity on social media. The paper is organized as followed. Section 2 discusses related work to this study. In section 3, the platform Instagram is introduced and image popularity is defined. Section 4 describes popularity factors that will be investigated. Section 5 describes the methodological approach and dataset collection characteristics. In section 6 the results of this study are presented. Section 7 describes the conclusion and ideas for future work.
2. RELATED WORK
A recent study [10] on Flickr shows certain objects in photos, such as bikinis, perfume and revolvers have a strong positive impact on popularity on this platform. The study also found social cues such as number of friends, tags, title length and description make good predictors for image popularity. A study [19] analyzing 8 million Instagram images on over 30 image features suggests there exist significant correlations between certain features (such as color hue, brightness and texture patterns) and image popularity. A study concerning content popularity on YouTube and Digg.com [9] suggests popularity growth can be predicted by a linear model, by measuring the first two hours of growth for stories on Digg and the number of views on YouTube, and then forecast their popularity for 30 days ahead with a relative error of 10%.
Studies addressing image aesthetics [11] [12] show certain visual properties, such as high photographic quality, make certain photos, in general, more aesthetically appealing than others. A study [14] that was done by using a large set of Flickr images
suggests correlations exist between positive sentiment and earthy and skin tone colors. Research addressing sentiment prediction in images [13] suggests sentiment for visual content can be predicted by a combination of textual metadata and feature analysis. A study addressing photo-popularity on Flickr [15] focuses on predicting popularity of images based on context, contextual information of the user who posted the image and contents of the image. The results show it is possible to predict image popularity, i.e., whether the images will receive high or low amount of comments on Flickr, in a cold start scenario (where no or little interaction data exists) with up to 76% accuracy. A study [16] investigating user preferences for visual summaries suggests the user’s criteria for selecting images for these summaries can be linked to aesthetic appeal and sentiment amongst other factors.
3. INSTAGRAM
Instagram is a social photo-sharing service, which allows users to upload captured images and videos and quickly edit these by applying one of about 30 filters. The platform is mainly centered on images [18], which can be shared with followers directly inside the application. In January 2014, Instagram had around 150 million active users per month who had shared a total of sixteen billion images. Due to the service’s growing popularity, companies are increasingly using Instagram to promote their products and reach their target audience [18]. Users interact on Instagram by liking images and/or leaving comments. An image is liked by double tapping on the image or by pressing the like button underneath the image. About 2.1 billion likes are given and received daily [24]. Users can use tags to describe their image by adding words behind the # symbol that they consider descriptive for their image. Images can contain up to 30 tags. Geo-tagging is done by enabling the phone’s geo-location and photomapping function inside the application. Users can follow each other to receive the followings’ images into their newsfeed.
4. IMAGE POPULARITY FACTORS
In this study, image popularity is measured by the amount of likes an image receives. Prior to this study, several factors are determined that are assumed to have potential impact on image popularity for this platform. The factors are generally based on previous studies and common intuition. A description and motivation per factor are further described below.
• Amount of followers
As stated in Introduction, popular content can become more popular by benefitting from the rich get richer dynamic. Therefore, it is likely that the amount of followers may have a positive influence on popularity. As the image receives more views, this might result in an increased number of ratings.
• Number of tags
The number of used tags may potentially influence image popularity. An image with more tags could generate more views, as it may be easier to retrieve, which makes it likely to receive an increased number of ratings.
• Content of tags
The content of tags may potentially influence image popularity, as certain tags are likely to be more used than others, and therefore generate more views.
• Uploaded time
It is assumed that at certain times of the day, online users are more active [18]. As views are required in order to receive response, the
time of the day content is uploaded (the uploaded time) may influence popularity.
• Filters
Previous research, such as [19] suggests certain visual features in Instagram images, such as blue hues, generate a higher number of likes than others. Thus, it seems likely certain filters may generate a higher number of likes as well.
• Semantic Concepts
The object depicted in the image may have influence on image popularity, as certain objects are likely to be considered more appealing or interesting than others.
5. METHODOLOGICAL APPROACH
The methodological approach for this study consists of a combination of an automatic analysis and two user studies. First, a preliminary user study is conducted to explore whether the hypothesis of factors earlier named match the responses and motivations of users on Instagram. The study consists of a questionnaire made with Google Survey Tools [20]. Thirty active Instagram users are recruited to fill out this questionnaire that concerns demographics, motivations to tag images, activity and motivations to like images. The results of this study are then used to determine which metadata to collect for the second part of the study. The second part of the study consists of an automatic analysis of a relatively large set of images. For this, a dataset is collected from Instagram in order to test which relations can be identified between popularity and the factors named in the section 4. The correlations are analyzed by using Python and statistics software SPSS. The dataset collection methodology is further described in section 5.1.
Furthermore, a main user study is conducted. This study consists of a questionnaire made with SurveyPlanet [21]. A relatively large group of participants are recruited to rate a set of 40 images from the earlier collected dataset, consisting out of 10 highly popular images, 10 popular images, 10 moderately popular images and 10 least popular images on Instagram. The images are presented to them in a random order and rated according to their image content. Lastly, the semantic concepts of the images are determined and the popularity on Instagram and in the user study is compared.
5.1 Dataset collection methodology
To retrieve a dataset consisting of images with diverse image content, the images were collected by their geo-location. The locations were selected according to the top 10 most instagrammed locations of 2013 [25], which was published on the Instagram blog in the beginning of 2014. To make sure each image has had an identical timespan to be viewed and rated, images were collected that were taken at all ten locations on the same day and uploaded in the same timeframe, which was Monday April 7th, 2014 between local times 9:00 AM to 12:00
AM (15 hours). The day of the week and timeframe were based on results of the preliminary user study (cf. Section 6.1) and ‘The Fortune 500 Instagram Report’ [18] that included earlier research showing most interaction on Instagram takes place on Mondays, and users are most active during daytime until about 12 AM. The dataset was collected using the Instagram API and diverse scripts written in Ruby and Python. The collected dataset contains 1590 images with their associated metadata.
6. RESULTS
complete study. First, an overview is given of the results obtained in the preliminary user study in section 6.1. Then, section 6.2 gives an overview of the results obtained in the automatic analysis. Section 6.3 presents and discusses the results found during the main user study.
6.1 Preliminary user study
The preliminary user study attempted to identify behavior of users on Instagram and which factors contribute to image popularity. This was done in order to test whether these motivations matched the factors in the hypothesis. For this, thirty active Instagram users were recruited via a personal network and via an online forum, and were asked to fill out 12 questions concerning demographics and their behavior and activity on Instagram. The majority of participants were female and in the ages of 18-35 years old. The results show most users make use of Instagram every day (cf. Figure 1), and do this mostly in the morning (73%), early evening (50%) and in the late evening (50%).
Figure 1: User activity on Instagram.
All participants said to use Instagram to view images (100%) and almost all participants (87%) also use Instagram to share images themselves. Three fourth (75%) of the users said to mainly view images in the application’s newsfeed. A smaller part said to also view photos they searched on by tag (cf. Figure 2).
Figure 2: Image viewing inside the application.
More than half of the users (57%) makes use of tags ‘Sometimes’ and does this mainly to describe the photo (63%). Some participants also do this to attract views (37%).
Motivations of users to like an image were determined by asking participants to choose from a predefined list, based on common intuition. The results show the most important reasons according to these users to like an image on Instagram are 1) the image was posted by a friend, 2) the image is aesthetically appealing and 3) the image is funny. Figure 3 shows the complete results. Results about activity were used to determine which timeframe should be extracted in the metadata collection for the automatic analysis. The results of tag-use were used to determine whether tags potentially may influence image popularity and therefore should be extracted. Results of motivations for image popularity were used to determine whether
Figure 3: Motivations of users to like images
results in the automatic analysis could be explained by answers of users.
6.2 Automatic analysis
For the automatic analysis, a set of 1590 collected images from the ten most instagrammed locations are used to analyze whether relations can be revealed between relevant factors identified in the previous sections and image popularity on Instagram. Table 1 shows the number of images per location in the collected dataset.
Table 1: Number of images collected per location
Location Name City, Country Number Images
Siam Paragon Bangkok, Thailand 378
Times Square New York, USA 349
Disneyland Anaheim, USA 141
Bellagio Fountains Las Vegas, USA 13
Disney World Orlando, USA 95
Staples Center Los Angeles, USA 23
Central Park New York, USA 112
Dodger Stadium Los Angeles, USA 92
Suvarnabhumi Airport Bangkok, Thailand 318
For those factors that contain variables that can be ranked, Spearman’s rank correlation coefficient (Spearman’s ρ) is computed. Table 2 shows the results of this computation per factor.
Table 2: Results for Spearman’s ρ for factors followers and number of tags.
Popularity and followers
The result for the followers an account has and the popularity of the image, shows Spearman’s ρ = 0.777, which indicates a strong positive correlation. This means the amount of likes increase when the amount of followers of the account increases. This partly corresponds with results in the preliminary user study, where users said to mainly view images in the application’s newsfeed, which consist of only images posted by users they choose to follow.
Popularity and number of tags
The results for the relation between the number of tags used for an image and the images’ popularity show Spearman’s ρ = 0.082, which indicates a slight positive correlation. This suggests the popularity of an image somewhat increases when the amount of tags increases. However, the number of tags in this dataset is unequally distributed (cf. Figure 4), which makes investigating a clear correlation between number of tags and popularity challenging.
Figure 4: Distribution number of tags and their frequency.
Popularity and Filters
Users can choose to edit their photos with about 30 filters. However, in this dataset, the most used filter is Normal, which means no filter was used, and more than half (55%) of the dataset consists of images that applied no filter. After this, filters with names Amaro, Valencia and Lo-fi were most used (Figure 5).
As Figure 5 shows, filters in this dataset are unequally distributed. As certain filters are not used more than once or twice, only the 10 most used filters are taken into account.
Figure 5: Distribution of filters.
Figure 6 shows a bar chart of the results. The results for this comparison show images with the filter Normal received the highest mean amount of likes on Instagram. After this, in descending order, the images with filters X-Pro II, Valencia, Sierra and Lo-fi received the highest mean amount of likes. Images that were edited with the filter Mayfair received the least amount of likes on average among the ten analyzed filters.
Figure 6: Ten most used filters and image mean popularity. Popularity and uploaded time
At certain times of the day, it is assumed that users may be more active than other times. Thus, it makes it likely that images posted at these times receive more views and therefore more ratings. To investigate whether a relation exists between the time of the day an image was uploaded and popularity, a few outliers were removed in order to get a clear picture. Figure 7 shows a
Factor Followers Number of tags
scatterplot of all images that received up to 500 likes and the local time they were uploaded.
As the dots show no clear peaks at any certain timeframes, it does not indicate a significant correlation between certain uploaded times and popularity.
Figure 7: Scatterplot upload time and popularity.
However, to investigate this further, linear regression (r2) is also computed. This method is used to predict a linear relation between two variables. The result shows r2 = 0.007, whichsuggests the amount of likes in this dataset canbe explained by their uploaded time for 0.07% of the images. This indicates as well no strong relation exists between time an image was uploaded and the popularity it received.
6.2.1 Datasets sorted by popularity
The popularity in the complete dataset is unequally distributed, which can be seen in Figure 8. Therefore, the popularity factors are also tested separately, by splitting the dataset into four smaller sets sorted by their popularity. Table 3 shows the characteristics of these datasets
Table 3: Datasets sorted by popularity and number of images.
First, Spearman’s ρ is computed between followers and amount of likes for all four datasets. Then, Spearman’s ρ computed for the number of followers and amount of likes for all four datasets. Table 4 shows the results for these computations.
In all datasets, Spearman’s ρ shows a strong positive relation between amount of likes and amount of followers again. This confirms the earlier findings for the complete dataset, that amount of followers have a strong impact on image popularity.
For the relation between number of tags used with the image and image popularity, the Popular and Moderately popular datasets show slight negative relation.
Figure 8: Distribution of image popularity complete dataset.
Table 4: Results Spearman’s ρ for factors followers and tags.
This would indicate the amount of likes for these datasets decrease when the amount of tags increases. However, in the Highly popular and Least popular sets, a positive correlation is found between number of tags and popularity. The Highly popular set consists of 24 images (cf. Table 3), which could make this result less reliable. However, Figure 9 shows in the Least Popular dataset consisting of 1105 images, the amount of likes seem to slightly go upwards when the number of tags increases to > 8.
Figure 9: Scatterplot number of tags and popularity for dataset with least popular images.
Dataset Amount of Likes Number Images
Highly Popular 500 + 24
Popular 151 - 500 133
Moderately popular 51 - 150 341
Least Popular 0 - 50 1105
Category Followers Number of tags
Highly Popular 0.420 0.261
Popular 0.390 -0.065
Moderately Popular 0.357 -0.078
Figure 10: Scatterplots showing uploaded time and popularity.
This may indicate that tag-use only has a positive impact on popularity for up to about fifty likes, as the impact of tags shows to become negative in the moderately popular and popular sets.
Popularity and uploaded time
To test whether a relation exists between image popularity and time that the image was uploaded in these dataset, a scatterplot was created for each dataset. Figure 10 shows the scatterplot for the Highest popular set in the upper left corner, the Moderatley popular set in the upper right corner, the Popular set in the bottom left corner and Least Popular set in the bottom right corner, with uploaded time and image popularity. The scatterplots do not show any clear signifiant correlations. However, to further investigate the relations here as well, linear regression (r2) is computed per dataset. The results are shown in table 5.
Table 5: Results r2 for factor uploaded time and popularity.
The results show for these datasets as well no significant relation is found between popularity and the local time images are uploaded. This might be explained by users’ answers in the preliminary study where the majority participants said to be active on Instagram every day, through the whole day, which may indicate they get to view and rate most images.
Popularity and content of tags
In the preliminary user study, a part of the users confirmed to search for images by tag. For this reason, it is investigated whether certain tags are associated with the (highly) popular images in the datasets. First, all tags that were used per dataset were automatically counted with Python. As there are no predefined tags on Instagram most tags are only used once. Therefore, only the 5 most used tags per dataset are taken into account. The highly popular set included no tags that were used more than once, which makes it unlikely these images became popular by certain tag-use. Figure 11 shows the results for the other sets.
In all sets, the same tag nyc was the one most frequently used. In the Popular and Moderately popular set, the tag dodger was second most frequently used. In the least popular set, the second and third most frequent used tags are newyork and timessquare. As the tag nyc is used most frequently in every dataset, it does not indicate to have any impact on popularity. The Moderately popular set and the Popular set both contain some images with the tags dodger, la and losangeles. These three tags are not found among the most frequent tags in the Least popular set. However, as these three tags only occur in less than 50 images in total, it is hard to say whether any significant correlation between these factors exists.
6.3 Main user study
The main user study was conducted in attempt to investigate whether image content may contribute or has any correlation to image popularity on Instagram. The main user study was utilized by recruiting a group of 77 participants via two online forums and a personal network. The group consisted of males (25%) and
females (75%), mainly in the age group of 18 – 35 years. This age group was selected purposely as it covers a large part of the target group of Instagram [17]. The participants were presented with a survey, consisting of a set of 40 images. The selection methodology of these images is further described in section 6.3.1.
6.3.1 Dataset image selection
To test whether a correlation may exist between certain semantic concepts in images and their popularity on Instagram, a set of images was selected from the dataset earlier obtained for the automatic analysis. The selection of images to use for the main user study was done by picking ten images from each of the earlier split dataset (Highly popular, Popular, Moderately popular, Least popular), that were sorted according to their popularity on Instagram. The characteristics of these sets have been earlier shown (cf. Table 3). The ten images per dataset were picked with a linear interval, starting from the image on the top with highest popularity, to the image on the bottom with the lowest popularity inside this dataset. For images that had received a similar amount of likes on Instagram, the images were selected by their geo-location. This was done in order to get a more equal distribution of image locations. Figure 12 shows the resulting distribution of images and their captured location.
6.3.2 User study
The user study was utilized by presenting the recruited participants with a survey made with SurveyPlanet [21], containing 40 selected images in a randomized order.
Figure 12: Distribution captured locations of images in main user study dataset.
Dataset r2
Highly popular 0.196
Popular 0.005
Moderately popular 0.010
They were asked to rate these images according to the image’s content. They could do this by either selecting a like or a skip button underneath the image, where skip means the image did not interest them and like means they liked the image.
The results (cf. figure 13 and 14) show a large difference exists in the preferences of the participants in the main user study, and the popularity ratings of the same images on Instagram. To investigate this difference further, the semantic concepts of the images were determined.
6.3.3 Semantic concepts of images
The semantic concept of the images were determined by selecting 3-5 concepts per image based on viewing the image and selecting concepts that are likely to describe the objects from the Large-Scale Concept Ontology for Multimedia [22], a multimedia concept lexicon. The images and these 3-5 concepts were then presented to 11 independent participants that were asked to select which concepts they considered descriptive for the image. For each image, the concepts that had an agreement by at least 9 out of the 11 participants were then used as the semantic concept that described the image. Figure 13 shows the distribution of images and their concepts.
Figure 13: Distribution of images and semantic concepts.
6.3.3 Semantic concepts and image popularity
The images and semantic concepts, their mean popularity on Instagram and their mean popularity in the user study are compared in figure 14 and 15.
The results show, the most popular images on Instagram mainly contain people (Faces, Male Person(s), Female Person(s) and Groups), Stadium, Baseball and images taken at or depicting Amusement Parks. The least popular images contain a Building, a Fountain and multiple Cityscapes.
In the user study, the most popular images contain multiple Cityscapes, an Animal, a Restaurant, a Fountain, Cartoon Animations and images taken at or depicting Amusement Parks. The least popular images contain people (Faces, Male Person(s), Female Person(s) and Groups), a Building and a Computer/television Screen. In the results, differences are seen between images that generate popularity on Instagram, and images rated highly by participants outside of Instagram. One explanation could be the fact that Instagram is a social platform,
Figure 14: Semantic concepts and mean popularity on Instagram with logarithmic scale.
Figure 15: Semantic concepts and mean popularity in the main user study.
which makes images become popular due to the social context of the image such as friends, family or celebrities being depicted in the images. As shown in the preliminary user study, participants responded their main reason to like images is that a friend posted the image. This indicates the person uploading the image may be of as much importance as what the image depicts. Also, as the majority of participants in the preliminary user study said to only view images inside their newsfeed, images posted outside of these followings are possibly not seen and therefore not rated, independent of what the image depicts.
7. CONCLUSION & FUTURE WORK
This study attempted to reveal a range of factors that were assumed to potentially influence image popularity on social photo-sharing platform Instagram. The most significant relation is found between image popularity and the number of followers of the account posting the image. The preliminary user study showed most participants only view images of users they choose to follow, which may explain why images posted by accounts with many followers receive a high number of likes. Furthermore, no strong relation is found between uploaded time of images and the popularity it receives. The number of tags used to describe the image showed a slight positive relation to popularity for the dataset that contained images with the least amount of likes. This suggests tag-use only generates a higher number of likes up to a certain limit. The relation between image filters and popularity showed images with no filter received the mean highest amount of likes. However, as filter use is unequally distributed in this dataset this could affect reliability. The main user study showed images depicting people and groups generated more likes on Instagram than in the user study. Cityscapes received the highest amount of likes in the user study, while generating the least number of likes on Instagram. This could suggest the social context of images on Instagram may be of more importance than e.g., “objective” image content.
For future work, a larger dataset could be collected to further investigate correlations between image popularity and filter use, semantic concepts and tag content, as the number of images in this dataset did not contain enough information for these factors to show convincing results. Also, in future research, certain factors such as amount of followers, could be eliminated first in order to measure whether significant evidence for correlations can be revealed between a separate factor and image popularity on Instagram.
ACKNOWLEDGMENTS
I would like to thank Stevan Rudinac for providing me with valuable feedback and good suggestions, which has been of great help during this study. I would like to thank Frank Nack for being my second reader, and all participants for their time and effort that have been part of one of the user studies.
REFERENCES
[1] S. Asur and B. A. Huberman, “Predicting the Future With Social Media” in Web Intell. Intell. Agent Technol., vol. 1, pp. 492–499, 2010.
[2] M. Cha, A. Mislove and K. P. Gummadi “A
measurement-driven analysis of information propagation in the flickr social network” in Proc. 18th Int. Conf. World wide web - WWW ’09, p. 721, 2009. [3] Y.-J. Park and A. Tuzhilin, “The long tail of
recommender systems and how to leverage it” in Proc. 2008 ACM Conf. Recomm. Syst. RecSys 08, p. 11, 2008. [4] F. Figueiredo, F. Benevenuto and J. M. Almeida “The
tube over time: characterizing popularity growth of youtube videos” in Proc. fourth ACM Int. Conf. Web search data Min. - WSDM ’11, p. 745, 2011.
[5] M. Cha, H. Kwak, P. Rodriguez, Y. Ahn and S. Moon “I Tube , You Tube , Everybody Tubes : Analyzing the World ’ s Largest User Generated Content Video System” in Proceeding IMC ’07 Proc. 7th ACM SIGCOMM Conf. Internet Meas., pp. 1–14, 2007.
[6] S. Asur, B. A. Huberman, G. Szabo and C. Wang “Trends in Social Media : Persistence and Decay” in ICWSM, 2011.
[7] D. A. Shamma, J. Yew, L. Kennedy and E. F. Churchill “Viral Actions : Predicting Video View Counts Using Synchronous Sharing Behaviors” in ICWSM, 2010. [8] M. Tsagkias, W. Weerkamp, and M. de Rijke, “News
Comments: Exploring, Modeling, and Online Prediction” in Adv. Inf. Retrieval, Proc., vol. 5993, pp. 191–203, 2010.
[9] G. Szabo and B. A. Huberman, “Predicting the popularity of online content” in Commun. ACM, vol. 53, no. 8, p. 80, Aug. 2010.
[10] A. Khosla, A. Das Sarma and R. Hamid “What Makes an Image Popular ?” International World Wide Web Conference in WWW, 2014.
[11] R. Datta, D. Joshi, J. Li and J. Z. Wang, “Studying Aesthetics in Photographic Images” in Computer Vision – ECCV 2006, pp. 288–301, 2006.
[12] C. Bauckhage and K. Kersting, “Can Computers Learn from the Aesthetic Wisdom of the Crowd?” in KI - Künstliche Intelligenz, vol. 27, no. 1, pp. 25–35, Dec. 2012.
[13] D. Borth, R. Ji, T. Chen, T. Breuel and S. Chang “Large-scale Visual Sentiment Ontology and Detectors Using Adjective Noun Pairs” in ACM MM, 2013.
[14] S. Siersdorfer, E. Minack, F. Deng, and J. Hare “Analyzing and Predicting Sentiment of Images on the Social Web” in 2010.
[15] P. J. McParlane, Y. Moshfeghi, and J. M. Jose, “‘Nobody comes here anymore, it’s too crowded’; Predicting Image Popularity on Flickr” in Proc. Int. Conf. Multimed. Retr. - ICMR ’14, pp. 385–391, 2014.
[16] S. Rudinac, M. Larson and A. Hanjalic “Learning Crowdsourced User Preferences for Visual
Summarization of Image Collections” IEEE Transactions on Multimedia, vol. 15, no. 6, pp. 1231–1243, 2013. [17] J. Brenner, “The Demographics of Social Media Users —
2012”, 2013. Available at:
http://www.pewinternet.org/2013/02/14/the-demographics-of-social-media-users-2012/ (Accessed: June 3, 2014)
[18] “The Fortune 500 Instagram Report, Trackmaeven”, 2013. Available at:
http://trackmaven.com/blog/2013/09/the-fortune-500-instagram-report/ (Accessed: June 12, 2014)
[19] “Study: 6 Image Qualities Which May Drive More Likes on Instagram” 2013. Available at:
http://blog.curalate.com/image-qualities-that-drive- likes-on-instagram (Accessed: January 29, 2014)
[20] “Google Survey Tools” Available at:
http://www.google.com/drive/ (Accessed: June 12, 2014) [21] “SurveyPlanet” Available at:
https://www.surveyplanet.com/ (Accessed: June 29, 2014)
[22] “LSCOM Lexicon Definitions and Annotations” Available at:
http://www.ee.columbia.edu/ln/dvmm/lscom/ (Accessed: July 12, 2014)
[23] “Youtube Statistics' Available at:
http://youtube.com/yt/press/nl/statistics.html (Accessed: July 4, 2014)
[24] “Instagram” Available at: http://instagram.com/press (Accessed: July 4, 2014)
[25] “Top 10 Most Instagrammed Locations of 2013” Available at:
http://blog.instagram.com/post/69877035043/top-locations-2013 (Accessed: June 2, 2014)
