Transcending Conventional Binary Labels: Revamping Knowledge Tracing with VAE-Generated Image Representation

Knowledge tracking serves as a data mining task aimed at predicting students’ learning performance. Nevertheless, without analyzing students’ answer contents, it is unjustifiable to represent learning records solely by two states of correct and incorrect answers for the knowledge tracking task. To tackle this issue, we incorporate image data to stand for the answer result labels. Even for the same answer result label, different vectors are employed for representation. Firstly, handwritten digit recognition is introduced as an auxiliary task, and the VAE network is utilized to generate handwritten digit images. Secondly, the handwritten digit images generated by the VAE are used to represent the vectors of answer results. Finally, the auxiliary task is combined with the knowledge tracking task for joint training. Experimental findings suggest the introduced images can better represent the complex states of learning results, thus enhancing the performance of the model. On four knowledge tracking datasets, the average AUC index increases by nearly 2.8%.