Learning Evidential Cognitive Diagnosis Networks Robust to Response Bias

Jiawei Liu, Jingyi Hou, Na Zhang, Zhijie Liu, Wei He

Abstract: As a basic task of the intelligent education system, cognitive diagnosis aims to diagnose the knowledge proficiency of students and capture the complex relationship between students and exercises. Benefiting from big data, deep learning show advantages in cognitive diagnosis tasks. However, the general deep learning-based methods are sensitive to noise, and response bias is an inevitable problem in real-world situations. To address this challenge, we propose the Evidential Cognitive Diagnosis Model (EvidentialCDM), which introduces the evidential deep learning to neural cognitive diagnostic frameworks for estimating the aleatoric and epistemic uncertainties as well as maintaining the predicting performance. In addition, this paper proposes a new dataset, named Uncertainty ASSIST (UncASSIST), in order to better deal with this problem. Experimental results show the effectiveness of our method on both the publicly available ASSIST and our proposed UncASSIST datasets.