Assessment of skills and adaptive learning for parametric exercises combining knowledge spaces and item response theory

作者：Munoz Merino Pedro J^*; Gonzalez Novillo Ruth; Delgado Kloos Carlos

来源：Applied Soft Computing, 2018, 68: 110-124.

DOI：10.1016/j.asoc.2018.03.045

• Item response theory
• Knowledge spaces
• Adaptive learning

Many computer systems implement different methods for the estimation of students' skills and adapt the generated exercises depending on such skills. Knowledge Spaces (KS) is a method for curriculum sequencing but fine-grained decisions for selecting next exercises among the candidates are not taken into account, which can be obtained with the application of techniques such as Item Response Theory (IRT). The combination of KS and IRT can bring advantages since the semantics of both models are included but some issues such as the required local independence of IRT should be considered. In addition, an open issue is how to handle with parametric exercises for skill modelling, i.e. exercises which are not static content but that can change from instance to instance depending on some parameters and a student can try to solve them again with different parameters after correct resolution. The correct inclusion of several instances of the parametric exercises on the adaptive decisions is important since the adaptation process can improve. This work describes two new algorithms for skill modelling and for adaptation of exercises that integrate IRT and KS to have a more powerful approach with more knowledge in the models and at the same time provides a solution for taking into account parametric exercises where a student should solve an exercise correctly several times to get proficiency. We have evaluated the different skill modelling algorithms using real data of students from their interactions in an Intelligent Tutoring System, and the correspondent adaptation algorithms using a simulator. Results show that the accuracy of the prediction is good with values of RMSE under 0.35. Both proposed algorithms got similar results on the accuracy of the prediction but one of them is better regarding performance. Changes of the buffer size for the MLE in IRT did not have a significant effect on the accuracy and on the performance. There is a tradeoff for selecting one of the two proposed algorithms: while the first algorithm has better performance time for the calculation of the ability (because there is no need of calculation of local abilities), the second algorithm has better performance time for the selection of the next exercise and better accuracy and depending on the scenario one or another should be selected.

• 出版日期2018-7