Predicting Multiple Risky Behaviors via Multimedia Content

Abstract

Risky behaviors pose a growing threat to our society. In the case of drug consumption, according to National Survey on Drug Use and Health, “substance abuse costs our society more than 484 billion dollars a year”, which is about thrice of what we spend on cancer. Researchers have started studying risky behaviors through big data from social media. However, to our best knowledge, most of the existing schemes focus on only one risky behavior, despite that the research in public health and psychology has shown us that there exist correlations among risky behaviors. In this work, in order to exploit such correlation, we select five risky behaviors, namely drug consumption, drinking, sleep disorder, depression, and eating disorder. Furthermore, we propose a deep learning neural network constructed by combining recurrent neural networks (RNN) and convolutional neural networks (CNN) to effectively predict whether an Instagram user will conduct which kind of risky behaviors in the near future.