Learning from Heterogeneous Student Behaviors for Multiple Prediction Tasks

Abstract

Prediction tasks about students have practical real-world significance at both student level and university level. For example, predicting if a student will fail to graduate can alert the university student affairs office to take predictive measures to help the student improve academic performance. In this paper, we focus on making multiple predictions together, since leaning the model for a specific task may have the data-sparsity problem. With the rapid development of smart campus, the university is accumulating a large amount of heterogeneous data of student behaviors, such as entering libraries behavior, entering dormitory behavior. In this paper, we propose to learn from heterogeneous student behaviors for making multiple predictions about students. However, leveraging heterogeneous behaviors have two main challenges. First, student profiles have a large impact on their behaviors and have not been well modeled in previous studies. Second, behaviors of different days will have different degrees of impact and should be treated unequally. To address these challenges, we propose a novel variant of LSTM and a novel attention mechanism. The proposed LSTM is able to learn student profile-aware representation from the heterogeneous behavior sequences. The proposed attention mechanism can dynamically learn the different importance degrees of different days for every student. With multi-task learning, we can deal with multiple perdition tasks at the same time to alleviate the data-sparsity problem. Qualitative and quantitative experiments on a real-world dataset collected at Shanghai Jiao Tong University (SJTU) have demonstrated the effectiveness of our model.