Knowledge-Guidance Based Multi-Relational Graph Spectral Clustering Using a Novel Structural Similarity Measure.

A group of objects along with multiple relations between them are usually modeled as a multi-relational graph, where multiple edges between vertices represent different relations between objects. The goal of multi-relational graph clustering is to partition the vertices in the graph into different clusters such that the vertices within the same cluster exhibit higher similarity than those across different clusters. Among the various existing multi-relational graph clustering methods, ensembling methods, though straightforward, fail to integrate all relations between vertices during the clustering process. To address this limitation and retain the intuitive nature of such methods, we propose a novel Knowledge-Guidance based Multi-Relational Graph Spectral Clustering (KG-MRGSC) algorithm. In this algorithm, the fusion of all relations between vertices is achieved at the level of partition matrices during the process of spectral clustering based on each relation. Specifically, a novel structural similarity measure is defined for constructing the similarity matrix when performing spectral clustering. This measure quantifies the similarity between pairwise vertices by considering various types of connectivity and highlighting the importance of the direct connection between them, which provides a reliable clustering basis for the proposed algorithm. To evaluate the quality of resultant clusters, two novel evaluation indices are designed based on the proposed structural similarities within clusters and across clusters. The validity of the proposed structural similarity measure and the superior performance of the proposed algorithm are demonstrated through comparisons with existing state-of-the-art methods in the experiments.