Prediction of collective actions using deep neural network and species competition model on social media

Abstract

Collective actions that can affect government management and public security (e.g., mass demonstrations), usually undergo long term development and originate from small and uncertain social media activities. Thus, researchers try to identify a collective action from various aspects such as changes in communication patterns, emerging keywords, and social emotions. Many studies aim to predict whether regular social media activities can evolve into collective actions, but the accuracy of these predictions is far from desirable. To address such a problem, we propose a framework named PFDNN which can predict the occurrence probability of collective actions every single day in the next month, so as to provide a reference for early decision-making. The framework consists of two parts: collective emotional contagion prediction and deep neural network with fully-connected layers (DNN) prediction. First, we implement the emotional contagion prediction based on species competition model to forecast user’s emotional state. Second, we model the DNN prediction as a binary classification problem that can be implemented using a DNN discriminator based on emotional contagion prediction. The DNN discriminator considers early premonitions based on the number of tweets, the embedded emotions and the number of violence-related words in the tweets during a specific timeframe, and automatically labels the early premonitions according to the number of reports published in the mainstream media. For evaluation purpose, we analyze the topics related to the “Arab Spring” from over 300,000 social media entries using TensorFlow. The results demonstrate that our prediction framework performs better than other representative methods.