A Multi-Task Knowledge-Tracking Model with a Novel Representative Approach to Problem Difficulty and Student Ability

Abstract

Question difficulty and student ability are important factors that affect students’ correct answers. Because existing knowledge-tracking models fail to consider these factors, they cannot accurately predict the results of students’ answers. In order to explore question difficulty and student ability information more accurately and to improve the accuracy of model prediction, this paper proposes a multi-task knowledge-tracking model (MTLKT) with a novel representative approach to question difficulty and student ability. The model first used the idea of multi-task learning to share underlying information and parameters, to jointly train, and to obtain an information difficulty representation vector consisting of skill difficulty and question difficulty. Then, combined with student learning process, a performance bias function was introduced to improve the attention mechanism and obtain a vector for student current knowledge state and a vector for question-solving performance, thus obtaining a vector for student ability information representation. Finally, the above vectors were concatenated and input into the model as a new representative embedding vector. The experimental results of three real-world data sets showed that our model had great improvement in the evaluation criteria of AUC and ACC and had a better predictive performance than the existing advanced knowledge-tracking models.

Keywords

• knowledge tracking
• multi-task learning
• attention mechanism
• education data mining
• deep learning