UNAGI: Unified neighbor-aware graph neural network for multi-view clustering

Abstract

Multi-view graph refining-based clustering (MGRC) methods aim to facilitate the clustering of data via Graph Neural Networks (GNNs) by learning optimal graphs that reflect the underlying topology of the data. However, current MGRC approaches are limited by their disjoint two-stage process, where the graph structure is learned in the first stage before the GNN messages are propagated in the subsequent stage. Additionally, current approaches neglect the importance of cross-view structural consistency and semantic-level information and only consider intra-view embeddings. To address these issues, we propose a Unified Neighbor-Aware Graph neural network for multi-vIew clustering (UNAGI). Specifically, we develop a novel framework that seamlessly merges the optimization of the graph topology and sample representations through a differentiable graph adapter, which enables a unified training paradigm. In addition, we propose a unique regularization to learn robust graphs and align the inter-view graph topology with the guidance of neighbor-aware pseudo-labels. Extensive experimental evaluation across seven datasets demonstrates UNAGI’s ability to achieve superior clustering performance.