Computerized Adaptive Testing Framework Based on Excitation Block and Gumbel-Softmax

Abstract

Computerized Adaptive Testing (CAT) is a personalized assessment method that adaptively selects the most suitable questions for students of different abilities based on their response data. Its primary goal is to effectively measure students' proficiency in a specific subject in a shorter period. The selection algorithm is pivotal in CAT. The current algorithms inadequately consider the impact of knowledge concept weights in question and student potential factors (e.g., memory) on question selection. In addition, most algorithms primarily focus on accurately predicting students' abilities, neglecting critical factors such as concept diversity and question exposure rate, which are essential for model effectiveness. Therefore, this paper introduces a new framework for CAT, GECAT. It proposes a selection algorithm based on an excitation block to learn the weights of each knowledge concept in the questions and analyze the impact of student potential factors on their answering performance, thereby selecting more suitable questions for students. Additionally, it views CAT as reinforcement learning, introducing Gumbel-Softmax to provide students with diverse, non-repetitive, and valuable test questions. The experimental results on three real-world datasets demonstrate that the proposed CAT framework improves ACC and AUC by 0.71% and 0.86%, respectively, while reducing question exposure rate and overlap rate by 1.33% and 1.59%, respectively.