Clustering Learning Objects Collections Using Cluster Ensembles

Learning Object Repositories are increasingly being used in learning systems to provide high-quality, reusable educational materials. A relevant data mining problem associated with the automatic categorization of learning objects is the discovery of intrinsic classes based on the textual contents of the meta-data records. In this paper, we present a cluster ensemble method, that is applicable for distributed clustering of learning objects, and for overcoming issues of large-sized collections and very high dimensionality of the vocabulary space. First, an initial significant reduction of the vocabulary space is proposed. Subsequently, a cluster ensemble method, consisting of three stages is proposed. Base clusterings for multiple random subsets of the data are generated, followed by the application of an adaptive voting algorithm for resolving the cluster label mapping problem. Finally, a consensus clustering is extracted by merging similar clusters that may be produced by fineresolution base clusterings, and maximum likelihood principle is applied on a resulting assignment probability matrix. The experimental analysis shows improvement in the consensus clustering quality over the base clusterings, and a competitive performance compared to clustering the original collection, as measured using normalized mutual information. The analysis validates the applicability of the proposed ensemble method.