||Predicting Latent Trends of Labels in the Social Media Using Infectious Capacity
Sonoda, YukiIkeda, Daisuke
International Journal of Future Computer and Communication
380 , 2015-12 , International Journal of Future Computer and Communication : IJFCC
This paper is devoted to predicting trends on the social media. Typical methods in the literature are based on temporal changes in usage of words or phrases on the media, and try to find a rapid increase, called a burst, of them. Therefore, these methods can be applied only after a burst is emerging. In this paper, we propose an index, called the infectious capacity, to detect potential trends on the social media before they would emerge. To achieve this, we focus on labels and items, and predict trends of a label, instead of those of a target object, such as contents of a social media, where an item is a concept represented by an object and a label categories items. On a photo sharing service, for example, a photo is an object, a tag is a label, and concepts represented by a photo are items for the photo. Using labels and items, the infectious capacity for a label is defined as the ratio of the variety of items with the label to the number of occurrences of the label in given data. That is, the larger value an infectious capacity of a label is, more infectious the label is. Our experiments on real data showed that the infectious capacities for most labels are substantially constant over time. This result means that we can forecast the variety per usage for a label just after the label is used. Moreover, we found that infectious capacities for popular labels have similar values. Combined with the first result, we are able to predict latent trends before labels become popular. In fact, this is also supported by experiments on tweets, where we were able to find potentially popular hashtags, regarding hashtags as labels, before they become popular. As far as the authors know, this is the first result of future trend prediction on the social media.