Exploring students' learning attributes in consecutive lessons to improve prediction performance

Building an understandable student prediction model has an essential role to play in the educational environment. Most current prediction models are difficult for teachers to interpret. This poses problems for model use (e.g., improve student performance, interventions and allow a feedback process). In this paper, we propose a new approach in building a practical model by identifying a number of attributes in comment data that reflect students' learning attitudes, tendencies and activities involved with the lesson. We check the capability of an attributes representation model compared to the Latent Dirichlet Allocation (LDA) model that represents student comments as a statistical latent class 'Topics.' In addition, we employ a Multi-Instance Learning (MIL) method to all available information for each student to improve the efficiency and effectiveness of classical representation for each lesson to predict final student performance. Computational experiments show that when the model is regarded as MIL, the prediction performance achieves better results than those based on single instance representation for each lesson.