Publication
Individualised Mathematical Task Recommendations through Intended Learning Outcomes and Reinforcement Learning
Alexander Pögelt; Katja Ihsberner; Norbert Pengel; Milos Kravcik; Martin Grüttmüller; Wolfram Hardt
In: Angelo Sifaleras; Fuhua Lin (Hrsg.). Generative Intelligence and Intelligent Tutoring Systems. International Conference on Intelligent Tutoring Systems (ITS-2024), located at 20th International Conference, ITS 2024, June 10-13, Thessaloniki, Greece, Pages 117-130, Lecture Notes in Computer Science (LNCS), Vol. 14798, ISBN 978-3-031-63027-9, 978-3-031-63028-6, Springer, Cham, 6/2024.
Abstract
Guiding students towards achieving the Intended Learning Outcomes (ILOs) of an academic module as part of a mentoring process presents a significant challenge, as it is important not only to emphasize the necessary skills, but also to consider the ongoing personal progress towards achieving a learning outcome. In addition, most educational content is presented in a ‘one-size-fits-all’ way, without taking into account the individual needs of students. In this paper we present a recommendation system based on Reinforcement Learning (RL) that derives its suggestions from the students’ progress towards achieving the ILOs and the current relevance of the ILOs, according to the specific didactic design of the module. The taxonomy model proposed by Anderson and Krathwohl, serves as the groundwork for abstracting ILO progress, temporal relevance, and the affiliation of recommendation items. In the process of creating a recommendation pool, experts identified the mathematical concept and the taxonomy level addressed by existing e-assessments in order to identify their possible association with ILOs. The RL agent utilizes this dynamic measurement of the student’s ILO progress - measured by the Bayesian knowledge tracing algorithm - to improve its recommendations, contributing to the ongoing personalisation of learning paths. In our evaluation, which utilized a test set of 129 mathematical tasks, the tested RL algorithms significantly outperformed a random baseline, underscoring the potential of this approach to enhance personalized learning within the realm of higher education mathematics.