Unsupervised Clustering of Skills for an Online Learning Platform

Afaf Ahmed, Imran A. Zualkernan, Hend Elghazaly

Abstract: Online learning platforms are generating an enormous amount of data that lends itself to unsupervised learning. This paper presents a case study where assessment data from two online platforms was used to cluster students into similar groups. The long-term objective of this research is to incorporate the clustering information into the personalization mechanisms. K-means was used to cluster students for 10 Skills. K-means was able to create a small number of clusters with reasonable internal validity with an average silhouette width of 0.32 (sd=0.05). The clusters were non-trivial as gender, school or class could not explain the clustering with an average Adjusted Rand Index (ARI) of 0.049 (sd=0.03). Most importantly, only a small subset (18%) of attempted questions could be used to explain accurately (Average F1-measure = 89.43) why the students were grouped into clusters. These keystone questions can be used to further enhance the personalization mechanisms.