Modeling question difficulty for unbiased cognitive diagnosis: A causal perspective

Abstract

Cognitive diagnosis is an intelligent education task that aims to learn students’ cognitive states on knowledge concepts based on historical answering logs over questions. Existing studies focus on modeling the interactions between students and questions through either manual-designed functions (e.g., logistic function) or complex neural network structures. However, such studies neglect the question difficulty bias, i.e., questions exhibit uneven distribution on the answering frequency, as simple questions are answered more times than difficult ones for the given concept. To tackle this issue, we present a Causal Cognitive Diagnosis Framework (CausalCDF), which considers the question difficulty bias and could be readily integrated with traditional diagnostic models for better cognitive diagnosis. Specifically, we first analyze the effect of question difficulty (acting as the confounder) on student performance via a causal graph. Then we eliminate the bad effect of the confounding difficulty bias via causal intervention in model training. We instantiate CausalCDF on five representative diagnostic models and perform extensive experiments on two real-world datasets. Empirical studies prove the effectiveness of CausalCDF compared to existing studies.