Graph-based Clustering Algorithms – A Review on Novel Approaches

Abstract

Classical clustering algorithms often require an a-priori number of expected clusters and the presence of all documents beforehand. From practical point of view, the use of these algorithms especially in more dynamic environments dealing with growing or shrinking corpora therefore is not applicable. Within the last years, graph-based representations of knowledge such as co-occurrence graphs of document corpora have gained attention from the scientific community. Accordingly, novel unsupervised and graph-based algorithms have been recently developed in order to group similar topics, represented by documents or terms, in clusters. The conducted work compares classical and novel graph-based algorithms, showing that classical clustering algorithms in general perform faster than graph-based clustering algorithms. Thus, the authors' focus is to show that the graph-based algorithms provide similar clustering results without requiring an hyperparamter k to be determined a-priori. It can be observed that the identified clusters exhibit an associative relationship reflecting the topical and sub-topical orientation. In addition, it is shown in a more in-depth investigation that the Seqclu (sequential clustering algorithm) can be optimized performance-wise without loss of clustering quality.