What's in a Caption?: Leveraging Caption Pattern for Predicting the Popularity of Social Media Posts

In the past few years, social media has become an integral part of modern society. It also has surfaced as an influential tool that helps a business or individual in gaining identity and reputation. Predicting the popularity of images before they are posted on social media thus may have a profound impact to reveal individual preference and public attention. However, an accurate prediction is a challenging task, mainly on account of factors that play a part in this. Previous studies, although achieve favourable results, overlook one unique characteristic of semantics in textual metadata, i.e., the language modeling, to better model the context information of a post. To that end, we propose to exploit the language modeling features together with user profile and post metadata features. The language model features are extracted by utilizing the probability of word occurrence, while the user profile and post metadata features are provided as attributes by the original data source. Several state-of-the-art statistical modeling techniques are employed to investigate the performance of the proposed features on different estimation procedures. Experiments on a large-scale Flickr dataset demonstrate the benefits of the proposed features on predicting the popularity of social media posts.