No abstract

No abstract

Computerized adaptive testing (CAT) targets to accurately assess the student’s proficiency in the required subject/area. The key issue is how to design a question selector that adaptively selects the best-suited questions for each student based on previous performance step by step. Most existing question selectors execute via greedy metric functions (e.g., question information and uncertainty), which can not effectively capture data characteristics. There also exist learning-based question selectors that redefine the CAT problem as a bilevel optimization problem, where the parameter learning of the question selector and the student proficiency estimation model are coupled, which is not flexible enough. To this end, in this paper, we propose a novel CAT framework with Decoupled Learning selector (DL-CAT). Specifically, we first use the currently estimated student ability and question characteristics as input and design a deep learning-based question selector to predict question selection scores. Then, to address the issue that there is no ground truth to measure the quality of the selected question, an approximate ground-truth and a pairwise rank loss function are specially designed to update the parameters of the question selector independently. Extensive experiments on two real datasets demonstrate that our proposed DL-CAT has certain advantages in effectiveness and significant advantages in efficiency.