Student Classroom Knowledge Tracking based on Deep Semantics-robust Network

Classroom knowledge tracking aims to predict future performance given past performance of students in educational applications. Although current classroom knowledge tracking methods achieve great prediction performance, there exist still two issues: (1) They exhibit sharp changes in knowledge state due to varying responses of students, resulting in semantic shifts in modelling knowledge state of students. (2) Transformer-based methods lack temporal information, limiting their ability to capture the gradual and cumulative nature of learning over time. To this end, a deep semantics-robust network is proposed via improving classroom knowledge tracking of students (DSN-KT) from three aspects, i.e., the data, the model, the loss, which significantly boosts the model stability and accuracy. Specifically, DSN-KT conducts the knowledge augmentation to generate data of different views for stable and robust estimation of knowledge states. And then, DSN-KT introduces the deep semantic learning within the transformer architecture with a time-cumulative attention, to capture the temporal dynamic information of students in learning knowledge. Meanwhile, DSN-KT devises the knowledge state prediction to provide the optimization function that optimizes prediction accuracy using semantic contrastive loss and cross-entropy loss. Three components work seamlessly to enable deep exploration and capture of students’ knowledge state patterns. Finally, extensive experiments on four datasets show superiority and effectiveness of DSN-KT.