Consensus representation-driven structured graph learning for multi-view clustering

Zhibin Gu, Songhe Feng, Jiazheng Yuan, Ximing Li

Abstract: Graph-based multi-view clustering has gained increasing attention due to its ability to effectively unveil complex nonlinear structures among data points from various views. Nevertheless, prior studies usually focus on amalgamating multiple similarity graphs derived from the initial data into a consensus one to serve the subsequent clustering task, leading to the clustering performance is notably contingent upon the inherent quality of the original features. Moreover, many prevailing approaches employ a two-phase methodology comprising graph construction followed by graph partitioning, which impedes the acquired graph from manifesting a structure conducive to the requirements of the clustering task. To overcome these issues, we propose an ONE-Step graph-based multi-view clustering via Early Fusion (ONESELF) method, which jointly conducts the robust latent representation extraction and the target structured graph construction into a cohesive optimization formulation. Specifically, a robust latent representation compatible across multiple views is firstly extracted from the original multiple features to mitigate the impact of inevitable noise and outliers. Subsequently, a consensus graph is formed by incorporating a connectivity constraint based on the latent representation, enabling direct extraction of clustering labels from the resulting graph without requiring further post-processing steps. Furthermore, we present an adept algorithm to efficiently optimize the objective function. Empirical results on seven benchmark datasets substantiate the superiority of our proposed method over several contemporary algorithms.