Time-and-Concept Enhanced Deep Multidimensional Item Response Theory for interpretable Knowledge Tracing

Abstract Knowledge Tracing (KT), namely tracking the knowledge conditions of each student across time, has always been challenging due to the latent and time-varying characteristics of knowledge states. Traditional psychometrical frameworks lack the ability to extract rich representations of exercises or examinees. Deep learning KT models have shown comparable performances, but uninterpretable model parameters have limited their applications. Furthermore, existing frameworks usually cannot handle temporal factors appropriately, as most of them simply apply stochastic processes to simulate fluctuations on knowledge states over time. In this paper, we propose a new framework named Time-and-Concept Enhanced Deep Multidimensional Item Response Theory (TC-MIRT) that integrates the parameters of a Multidimensional Item Response Theory into an improved recurrent neural network. Specifically, two enhanced components are constructed to empower our model with the ability to perform trend forecasting and to generate explainable parameters in each specific domain of knowledge. Experiments implemented on two real-world datasets show that our framework outperforms state-of-the-art KT approaches on performance prediction tasks. Moreover, extensive case analyses also demonstrate that the interpretable parameters of TC-MIRT can be used to evaluate the strengths and weaknesses of students.