Self-attention and forgetting fusion knowledge tracking algorithm

Abstract

Knowledge tracking is a method to determine students’ potential knowledge states based on their historical learning trajectory and to track students’ knowledge states in real-time to foretell their future learning circumstances. To solve the problem that existing algorithms ignore weak analysis of feature independence between different test questions, we propose the algorithm SATFKT. SATFKT integrates difficulty division and discrimination of test questions into input features. This helps knowledge tracking tasks better model and analyze students’ mastery of knowledge concepts. An algorithm, SAFFKT, is also proposed to further solve the problem of existing algorithms neglecting students’ memory and forgetting behavior. SAFFKT is formed by adding a forgotten update layer and a memory reading layer to the original hierarchical structure of SATFKT. The forgotten update layer helps the knowledge tracking task process the forgetting before the students answer questions. After students have finished answering questions, the knowledge state is further updated by the memory reading layer based on the answers. These two additional network modules help the model simulate students’ learning and forgetting behaviors more realistically. Compared with traditional knowledge tracking models and other models that consider forgetting behavior, our model predicts higher AUC values and proves the effectiveness of SAFFKT.