A Multitask GRU Deep Learning Model for Early Stage Prediction of Spatial Spread

Online social networks generate a huge amount of data of various types such as hashtags, images, and posts, which spread among users by several means such as re-posts, shares, likes, etc. Understanding of spreading phenomena has vital roles in the prediction and manipulation of data dissemination over social media as well as in real-world scenarios. There are several applications such as marketing, epidemiology, awareness campaigns, reportage, and virality which have a requirement of studying spreading phenomena. Past studies related to spreading revolve around volume and temporal dimensions of spread, such as predicting total infected users, predicting popularity (instances of a particular event/hashtags), and predicting the time when content receives a threshold number of infected users. However, with the availability of geo-tagged social network data, we have one more dimension of spread, i.e., the spatial dimension. In particular, adding the dimension of geographical space aids in further enhancement of dependent applications of location-based services such as region/target-specific marketing, advertisement, dialectology, and online campaigns. In this paper, we attempt to do early-stage prediction of the spread of hashtags (social media textual entities) across geographical space. To achieve this, we deploy a GRU (Gated Recurrent Unit)-based multi-task learning neural network model. Experimenting on three datasets from X (earlier known as Twitter) shows a significant improvement over the state-of-the-art.