Prediction modeling using Big data can be considerably enhanced by using network science along with machine learning informed by social science theories. In this research, we propose a big data approach to formulating a predictive model by integrating measures from networks of social interaction gleaned from large spatio temporal datasets. The prediction model goal is to identify students at risk of dropping out in a proactive and timely manner. We use a combination of commonly available (student demographic and academic) data in academic institutions augmented by implicit social interaction measures derived from students’ university smart card transactions. Furthermore, we develop a sequence learning method to infer students’ patterns of activities from their location check-ins. Since student retention data is highly imbalanced, we build a new ensemble machine learning classifier to predict students at-risk of dropping out. For model evaluation, we use real-world data on smart card transactions as well as other types of student information from a large educational institution. The experimental results show that the addition of campus integration and social interaction features refined using the ensemble method significantly improve prediction accuracy and recall. Zoom: https://arizona.zoom.us/j/183045568 (begins at 12:30pm AZ Time - please do not log on before that time)
