LinkFCM: Relation integrated fuzzy c-means

Most existing fuzzy clustering approaches group objects in a dataset based on either a feature-vector representation of each object, or pairwise relationship representation between each pair of objects. However, when both forms of data representations from different descriptions are available for a given dataset, we believe that a dual and cooperative analysis of feature-vectors (vector data) and pair-wise relationships (relational data) is likely to gain a more comprehensive understanding on the characteristics of the dataset, based on which a better clustering result may be achieved. In this paper, we develop a new fuzzy clustering approach called LinkFCM, which integrates pair-wise relationships into fuzzy c-means vector data clustering. The objective function of LinkFCM consists of two different ways to measure the compactness of clusters with respect to vector data and relational data, respectively, so that clusters are formed by utilizing these two forms of data descriptions. Our experimental study shows that LinkFCM is able to produce good clustering results for real-world document datasets by effectively making use of both content of documents and links among documents. This demonstrates the great potential of the proposed approach for data clustering, where pair-wise relationships are available together with features that describe each object.

► LinkFCM is formulated to cluster objects by making use of information from both vector data and relational data. ► LinkFCM provides the mechanism to integrate relational data into the popular fuzzy c-means clustering. ► Experimental studies on document datasets show that LinkFCM outperforms existing fuzzy approaches.