International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Extracting structural clusters from NMF feature matrix using Cosine Similarity-Based Weighted Voting

Authors :

Mehdi Rahimi¹ Keyhan Khamforoosh² Vafa Maihami³

1- Department of Computer Engineering, Sa.C. Islamic Azad University, Sanandaj, Iran 2- Department of Computer Engineering, Sa.C. Islamic Azad University, Sanandaj, Iran 3- Department of Computer Engineering, Sa.C. Islamic Azad University, Sanandaj, Iran

Keywords :

Non-negative Matrix Factorization (NMF)،Clustering،Community Detection،Cosine Similarity

Abstract :

In this paper, a novel post-processing method is proposed to enhance the stability and accuracy of clustering results obtained from Non-negative Matrix Factorization (NMF) on graphs. Unlike conventional methods that assign nodes to clusters directly based on the maximum value in each node’s feature vector, the proposed method employs a weighted voting mechanism based on cosine similarity. In this mechanism, the cluster of each node is determined through weighted voting by its neighbors, where the weight of each vote corresponds to the cosine similarity between the neighbor and the target node. The performance of the proposed method is evaluated on four real-world datasets as well as the synthetic Girvan–Newman dataset, and compared against the baseline method. Experimental results demonstrate that this approach leads to significant improvements in the clustering accuracy metric (NMI) in most cases. Moreover, an increase in modularity is observed across all datasets, indicating stronger intra-cluster cohesion from the graph’s topological perspective. This improvement is especially important in networks with unclear boundaries between clusters, because in such situations, taking advantage of the structural similarity between nodes and their neighbors can play an effective role in making more accurate decisions for cluster assignment.