Graph clustering using k-Neighbourhood Attribute Structural similarity

• Novel graph clustering algorithms (kNAS) is proposed, for overlapping community detection in large graph by combining the topological and attribute similarity that partitions the large graph into m clusters having high intracluster and low intercluster similarity.
• The core node in the graph is identified using Local Outlier Factor. Structural Similarity is based on grouping of nodes based on the neighbourhood of the core node and Attribute Similarity is achieved using Similarity Score.
• An objective function is defined for the faster convergence of the clustering algorithm.
• Density and Tanimoto coefficient are the validation measures used to define the effectiveness and quality of the proposed algorithm with the existing algorithms.

A simple and novel approach to identify the clusters based on structural and attribute similarity in graph network is proposed which is a fundamental task in community detection. We identify the dense nodes using Local Outlier Factor (LOF) approach that measures the degree of outlierness, forms a basic intuition for generating the initial core nodes for the clusters. Structural Similarity is identified using k-neighbourhood and Attribute similarity is estimated through Similarity Score among the nodes in the group of structural clusters. An objective function is defined to have quick convergence in the proposed algorithm. Through extensive experiments on dataset (DBLP) with varying sizes, we demonstrate the effectiveness and efficiency of our proposed algorithm k-Neighbourhood Attribute Structural (kNAS) over state-of-the-art methods which attempt to partition the graph based on structural and attribute similarity in field of community detection. Additionally, we find the qualitative and quantitative benefit of combining both the similarities in graph.

Figure optionsDownload as PowerPoint slide