Anti-collusion Dynamic Distanced Online Testing

Abstract

In this paper, we address the problem of minimizing collusion gain, the increase in students' test scores when they cheat by colluding with each other, in distanced online testing (DOT). Recent works seek to address this by requiring students to answer questions sequentially within designated time slots synchronized over all students and assigning questions from a large question bank in a manner that eliminates the possibility of a lower-competence student colluding with and copying from a higher-competence student by assigning the lower-competence student with any common questions before the higher-competence student. However, existing techniques require estimates of students' competencies and their collusion network as input. This poses a significant challenge to their widespread adoption in the real world, where such estimates are either unavailable or, at best, noisy. To address this challenge, we provide the dynamic DOT framework, where tests are conducted in multiple rounds and questions are assigned dynamically in each round using updated estimates of students' competencies and collusion network learned from students' performances in previous rounds. We show through extensive experiments on simulated tests that in the common regime, where the size of the question bank is close to the desired length of the test (number of questions assigned to each student), our dynamic DOT framework incurs significantly lower total collusion gain on average compared to tests conducted using both randomized and existing static question assignment schemes that do not update estimates of students' competence and collusion network.