Towards the Identifiability and Explainability for Personalized Learner Modeling: an Inductive Paradigm

Jiatong Li,Qi Liu,Fei Wang,Jiayu Liu,Zhenya Huang,Fangzhou Yao,Linbo Zhu,Yu Su

DOI: https://doi.org/10.1145/3589334.3645437

2024-01-01

Abstract:Personalized learner modeling using cognitive diagnosis (CD), which aims tomodel learners' cognitive states by diagnosing learner traits from behavioraldata, is a fundamental yet significant task in many web learning services.Existing cognitive diagnosis models (CDMs) follow the proficiency-responseparadigm that views learner traits and question parameters as trainableembeddings and learns them through learner performance prediction. However, wenotice that this paradigm leads to the inevitable non-identifiability andexplainability overfitting problem, which is harmful to the quantification oflearners' cognitive states and the quality of web learning services. To addressthese problems, we propose an identifiable cognitive diagnosis framework(ID-CDF) based on a novel response-proficiency-response paradigm inspired byencoder-decoder models. Specifically, we first devise the diagnostic module ofID-CDF, which leverages inductive learning to eliminate randomness inoptimization to guarantee identifiability and captures the monotonicity betweenoverall response data distribution and cognitive states to preventexplainability overfitting. Next, we propose a flexible predictive module forID-CDF to ensure diagnosis preciseness. We further present an implementation ofID-CDF, i.e., ID-CDM, to illustrate its usability. Extensive experiments onfour real-world datasets with different characteristics demonstrate that ID-CDFcan effectively address the problems without loss of diagnosis preciseness.