Investigation of Step Granularity for Adaptive Learning Strategies

Abstract: Intelligent tutoring systems (ITS) are softwares attempting to emulate human tutors. They do so by offering sequences of exercises for the students to train their newly-learnt skills, as well as to evaluate their competences. In order to estimate the knowledge mastery of each of their students, they make use of a student model. The purpose of student models is to assess a student’s knowledge based on observable features such as performances. Through this information,ITS can optimise the learning sequence of their users which means that fast learners will be recommended challenging questions while slower users will receive simpler problems. In classic configurations, each recommended exercise contains one single question, is labeled with a competence, and is taken into account into the internal student model right after its completion. This structure does not allow for the system to understand what part of the question, and thus what area of the competence was not mastered when the student fails to complete it correctly. In this project however, exercises labelled with one competence can contain different questions labelled with various subskills and diffculty levels, which all need to be submitted before the ITS updates its internal knowledge. This allow the algorithms to pinpoint where the weaknesses of the students lay with more precision. To this purpose, Skill-PFA, Diffskillty-PFA, Skill-BKT and Diffskillty-BKT, variants of Performance Factor Analysis (PFA) and Bayesian Knowledge Tracing (BKT) respectively have been developed. Those four variants are able to deal with exercises containing different questions, on top of handling more granular labels on the knowledge estimations. In our experiments, we show that our PFA adaptations are much more robusts and produces better results than any BKT variants. Furthermore, PFA’s biggest liability lays in parameters we can control, such as the training data size and the potentially mislabelled items. Additionally, we demonstrate that the models built for this project, Skill-PFA and Diffskillty-PFA consistently perform better than PFA. Indeed, increasing the granularity of the labels enables those new models to pinpoint more precisely where the students’ weaknesses and strengths lay