Item Deep Response Model: Applicable as both Measurement Model and Knowledge Tracing Model

Abstract

Knowledge tracing (KT), the task of tracking the knowledge state of each student over time, has been studied actively by artificial intelligence researchers. Recent reports describe that deep-IRT, which combines Item Response Theory (IRT) with a deep learning model, provides superior performance. However, its interpretability and applicability remain limited compared to those of IRT because item and ability parameter estimates depend on the order of the presented items. To overcome those difficulties, this study proposes Item Deep Response Model (IDRM), which models a student's deep response to an item by two independent networks. Experiments reveal that IDRM resolves difficulties of earlier models and increases predictive accuracies.