Clustering Educational Items from Response Data using Penalized Pearson coefficient and deep autoencoders

Abstract

Educational data mining techniques are very useful to analyze learner performance in purpose to optimize the approach of item-to-skill mapping. Therefore computing a degree of similarity between items using different measures based on the performance of the learner toward items, enhance the clustering of different items into knowledge components. This paper proposes a computational framework to group the elements of the corresponding knowledge component. The first phase of the framework represents a variation of Pearson coefficient to measure item similarity by applying a penalty score that is calculated from the number of hints taken by the learner during solving two items. The second phase applies a dimensionality reduction using deep auto encoders to improve the clustering accuracy. The experimental results show that clustering based on the penalized Pearson coefficient and the deep dimensionality reduction (PPC+DDR) outperforms basic clustering based on different similarity methods , with approximately +0.2 in Mean silhouette coefficient.