Efficiency and Stability of Clustering Algorithms for Linked Data

We are interested in finding clusters (“communities”) in networks of linked data, such as citation networks or web pages. Hierarchical clustering for networks is reviewed and an algorithmic improvement that leads to a significant performance increase is introduced. Our main focus is on the development of partitioning clustering algorithms that can deal with data represented only by link information (e.g., documents represented only by their citations) and the development of an EM algorithm for such data. A desirable property of clustering is stability; that is, small changes to the data should not lead to dramatically different clusterings. In our experiments with citation data we compare the hierarchical and partitioning clustering algorithms in terms of efficiency, stability and intra-cluster similarity. The problem of mining linked data (e.g., [4]) has become quite important as more and more information such as scientific publications or simple web pages is made available online. The most popular link mining tasks concentrate on finding communities in citation data [9] or in web pages from link topology [5]. Identification of terrorist networks [1; 8] and of fraud in telecommunication networks [2] are among the relevant applications which motivate research in this field. Clustering is an elementary data analysis step that is well examined for traditional machine learning settings and is now also being applied to linked data. When analysing linked data, it seems obvious to represent each node of this network either by its inbound, by its outbound or by both kinds of links. One can distinguish hierarchical agglomerative [6] and flat, partitioning clustering algorithms (e.g., [3]). Hierarchical clustering algorithms require a distance metric between pairs of instances to be defined, whereas k-means and EM with mixture models require the instances to be represented as a vector in feature space. Up until now, in most cases an agglomerative clustering algorithm was employed when clustering linked data. Yet this kind of algorithm is known to be very time consuming when clustering large numbers of objects and has shown to be sensitive to perturbations of the data to cluster [7]. We examine the applicability of partitioning algorithms to linked data and compare their performance in terms of efficiency, stability and intra-cluster similarity to the agglomerative algorithm. Our contribution is threefold. Firstly, we propose a caching strategy for hierarchical agglomerative clustering that improves its performance for clustering link data. Secondly, we derive an EM clustering algorithm for link data. Thirdly, we compare the clustering algorithms in terms of efficiency, stability and intra-cluster similarity for a publication database.