Enhanced Transformer: Knowledge Tracing with Incorporation of Temporal Features
Akarsh R Hegde, Sharma Akash, C H Adithya, B V Vaishnavi, Chetana Hegde · 2024
Knowledge Tracing (KT) plays a pivotal role in understanding and predicting students’ progress in the realm of education. Efficiently and accurately predicting a student’s likeliness of providing a correct answer to the upcoming questions and estimating their knowledge state are essential in the field of KT. However, existing research often overlook critical elements, such as students’ forgetting behavior and temporal aspects. Therefore to overcome these shortcomings, we introduce a novel Transformer-based architecture designed to predict students’ next answers with precision and estimate their knowledge state more effectively. We incorporate temporal features, including response time (the duration students take to respond), lag time (the time gap between successive learning activities), and the length of time students dedicate to interacting with video content and explanatory lectures. We have implemented our model on the EdNet KT3 dataset, and the results of our evaluation demonstrate remarkable accuracy with an improvement in ACC score of 17.80% and an AUC score of 11.14% compared to the baseline models we have mentioned below in the paper. Throughout our research, we explore key features that play a critical role in Knowledge Tracing framework.