Neural attentive knowledge tracing model for student performance prediction

A large number of anonymous log files are collected from the online education platform, and it is of great educational significance to use efficient algorithms for mining student's characteristics and predicting student's performance. To the best of our knowledge, existing models lack attention to the long-term performance of students. The interpretability of the operating results is weak. In addition, these models simplify the tracking of student knowledge points and are essentially unable to capture the relationship between skills in multi-skill exercises. We propose a new model, NAKTM, which divides user features into long-term and short-term features, and uses both to comprehensively express student abilities. At the same time, it uses the skills involved in the exercises as much as possible to jointly represent the characteristics of the exercises. Finally, we use the bilinear matching scheme in the hidden space to calculate the similarity between the students' ability and the exercises, and finally directly predict the learner's performance at the exercise level at the next moment. The experiment shows that our model achieves good experimental results without special processing of datasets.