Student Academic Performance Prediction Using Deep Multi-source Behavior Sequential Network

Online education is becoming increasingly popular and often combined with traditional place-based study to improve learning efficiency for university students. Since students have left a large amount of online learning data, it provides an effective way to predict students’ academic performance and enable pre-intervention for at-risk students. Current data sources used to predict students’ performance are limited to data just from the corresponding learning platform, from which only learning behaviors on that course can be observed. However, students’ academic performance will be related to other behavioral factors, especially the patterns of using Internet. In this paper, we utilize two types of datasets from 505 university students, i.e., online learning records for a project-based course, and network logs of university campus network. A deep learning framework: Sequential Prediction based on Deep Network (SPDN) is proposed to predict students’ performance in the course. SPDN models students’ online behavioral sequences by utilizing multi-source fusion CNN technique, and incorporates static information based on bidirectional LSTM. Experiments demonstrate that the proposed SPDN model outperforms the baselines and has a significant improvement on early-warning. Furthermore, it can be learned that Internet access patterns even have a greater impact on students’ academic performance than online learning activities.