One-Step Graph Fusion Fuzzy Clustering Network

Graph clustering plays a crucial role in uncovering implicit information within data which can be used to rationally classify potential data samples in unsupervised scenarios. In recent years, investigators have essentially focused on deep graph clustering, where deep learning techniques are effectively utilized to transform nodes in the graphs into representative vectors for subsequent clustering. Nevertheless, the existence of two main constraints within current methodologies warrants their their resolutions: (i) the majority of graph clustering methodologies are bifurcated into two-step, potentially culminating in clustering outcomes that deviate unpredictably from the actual truths; (ii) the prevailing graph clustering techniques chiefly employ hard clustering paradigms, that exhibit a lack of interpretability concerning data points situated in boundary regions that may pertain to multiple clusters. To address these challenges, we herein introduce a one-step graph fusion fuzzy clustering network (GFFCN) that incorporates fuzzy clustering into the graph structure by modulating its fuzziness coefficient to reveal the degree of membership of each data point to various clusters. Additionally, an expanded mutual information regularization component is incorporated to enhance its equilibrium and precision, and to introduce additional clustering layers that fortify the overarching network. Collectively, the GFFCN adeptly surmounts the constraints of the conventional two-step clustering methodology, facilitating the entire clustering framework to yield outcomes in a seamless end-to-end fashion, which in turn diminishes the discrepancies associated with the bifurcated clustering process. The extensive experiments conducted demonstrate that the proposed approach achieves state-of-the-art results in six data sets and effectively enhances the determinism and balance of clustering outcomes.