Image popularity for a user in social media can be attributed to variety of factors of which both image content as well as the social aspect plays significant role. With the explosion in number of photos clicked using Smartphones, it is important that the user can quickly search and find photos he may be interested in. According to report from KPCB analyst Mary Meeker in 2014, we now upload and share over 1.8 billion photos each day through various channels and social media. Since the numbers of people who have access to these images are also very large, getting one’s image viewed or liked by many people gives a feeling of being popular in one’s social circle and in turn an instant gratification. This makes image popularity an important driving factor for choosing and sharing the right image from a large collection of images. In this research we investigate what images are likely to be more popular so that a recommendation channel can be generated to allow the user to quickly access these photos.

Even though popularity of text (such as tweets) [1] [2] and videos (such as YouTube videos) [3]-[5] have been studied in recent years, image popularity prediction still remains a difficult problem. In our study, we have found that the most relevant research work in image popularity prediction [6] has been done using the Flickr dataset. This research is however a combination of both social aspects and image content. Also, the definition of popularity is number of views in Flickr. As opposed to these ideas, in our work, we address two issues in particular: First, we investigate whether there is a more relevant definition of image popularity and answer this question affirmatively. Second, we show that it is possible to predict image popularity with a significant accuracy based on image content alone. Although our definition of popularity is based on number of likes in Facebook, we have only used this information to label the training set. To the best of our knowledge, this work is the first to predict image popularity based on image content alone.

As already discussed, the most significant work on image popularity prediction has been done by Khosla et al. [6]. For other related work on popularity, there has been some work by Figueiredo et al. in video popularity prediction [7] [8]. However, their work is mostly dependent on social aspect like comments and tags associated with videos. Popularity prediction has been studied in other areas like popularity of online articles [9] and popularity of social marketing messages [10]. However, the study done in these papers comes under the category of predicting text popularity such as tweets or Facebook posts and predicting popularity of web pages. None of these in any way relate to the image popularity we discuss in this work. Xinran et al. in [11] have studied the prediction of clicks on Ads in Facebook as a classification problem. Although number of clicks associated with an ad can be considered as a form of popularity, click prediction of an Ad is not directly related to popularity.

In another interesting work by Justin Cheng et al. in [12] they study the prediction of cascading in social media such as Twitter and Facebook. Cascading refers to sharing and re-sharing posts in Facebook and re-tweeting tweets in Twitter. These sharing may include anything from text to images that is sharable through these channels. The growth of this cascade can also in some ways be considered as popularity of the sharable items. This however does not directly relate to our consideration of image popularity for individual users in social media particularly for two reasons. First, most of these sharable items consist of items that are strictly public and not private images from users. Second, along with content, there is a strict dependency on other social features that cannot easily be avoided. Although a separate study of popularity prediction for these sharable items is possible, in our context, due to their varied properties we have decided not to mix them up.

The definition of popularity by Khosla et al. in [6] is number of views in Flickr. In their work they have considered the normalized view count as the dependent variable for regression and the performance measure is in terms of rank correlation. Table 1 summarizes their results.

From Table 1, one can see that the contribution to prediction by image content alone is very low and most of the contribution is from social aspect. Another issue is the number of features used in [6] is very large for image content alone and yet contribution to prediction from all these features is very little. Table 2 summarizes this issue.

Each of these categories in image content (like Gist, Texture etc.) consists of a number of features ranging from 512 to 10,752. Computing all these features from a collection of images may not be a feasible scenario in devices such as Smartphones where one is limited in terms of memory and processing capabilities. Also in [6] the final prediction is number of views an image is likely to receive as a result of prediction but the conclusion whether or not the image is popular is a matter of subjective judgment (one may decide popularity based on number of predicted view count). In our work we look at this problem in slightly different way.

For a given set of feature vector X and binary label set B Bayes theorem can be stated as:

where, is the conditional density function. In our Naïve Bayes implementation in the previous section, the assumption is that

In order to verify this we plot the density function of one of the features Blur in Figure 6.

For a lot of features in our case the Normal distribution turns out to be a not so good fit. In our experiment we also observe that for certain features a Gamma or a Weibull distribution fits better than Normal distribution. Table 4 summarizes this observation.

We measure goodness of fit in terms of p-value greater than 0.05. From the summary in Table 4 we can see that some features even though fit well with Normal distribution, in other cases Weibull or Gamma distribution scores higher in terms of goodness of fit (Figure 6 and Figure 7).

We finally use this variable density function based Naïve Bayes to predict popularity. Figure 8 compares the precision and recall of Naïve Bayes with normal distribution (NBND) vs. Naïve Bayes with variable density function (NBVDF). With this modification we have been able to achieve a 52% precision and 80% recall.

A regression based popularity prediction has constrains such as: low accuracy when only image content is used, system overhead due to large number of features and highly dependent on social data which may not always be available. In most cases one is concerned with the final decision on whether or not an image is going to be popular and not be concerned about the number of views or likes. Thus solving the classification problem is much more intuitive and a simple Naïve Bayes variant works well without much system overhead. In future we plan to add contextual information to improve the overall precision and recall. In our work we have been able to show that one may choose to ignore any social aspect associated with the image and still achieve significant precision and recall to predict popularity of an image purely based on image features.

Arunabha Choudhury,Sriram Nagaswamy, (2015) A Bayesian Approach to Identify Photos Likely to Be More Popular in Social Media. Journal of Computer and Communications,03,198-204. doi: 10.4236/jcc.2015.311031