Mining Large-Scale News Articles for Predicting Forced Migration via Machine Learning Techniques

Many people are being displaced every day from all around the globe. Many of them are forced to leave their homes because of socio-political conflicts, human-made or natural disasters. In order to develop an early warning system for forced migration in the context of humanitarian crisis, it is essential to study the factors that cause forced migration, and build a model to predict the future number of displaced people. In this research, we focus on studying forced migration due to socio-political conflicts for which violence is the main reason. In particular, we investigate whether the degree of violence in a specific region can be detected from news articles related to that region and whether the detected violence scores can be used to improve the prediction accuracy. We investigate three techniques to extract the degree of violence from a corpus of news articles: ED-FE, TD-FE and SWSW. SWSW measures the semantic similarity between documents and a set of seed-words representing violence. ED-FE extracts violent events from news articles, which are the incidents related to attacks or the ones resulting in casualties. TD-FE uses topic modeling techniques to reduce the size of the information for easier analysis and filtering the violent incidents. Experiments indicate that ED-FE and TD-FE provide accurate violence scores which are very effective features for making forced displacement forecasts and using them in prediction models improves the prediction accuracy.