Graph Autoencoder Subspace Clustering

Abstract

Abstract: Most subspace clustering networks in current literature adopt autoencoders as the overall architecture, using convolutional layers within the autoencoder network to capture features of data points. However, shallow convolutional layers have limited feature extraction ability, as they fail to consider the overall image region. On the other hand, deep convolutional networks introduce difficulties in parameter tuning and require significant computational resources for training. Additionally, these features do not take into account the structural relationships among data points, thereby impacting the quality of features extracted by the subspace network.To address these issues, we propose a novel Graph Autoencoder Subspace Clustering (GASC) model. Specifically, instead of traditional autoencoders, we replace them with graph autoencoders, where the encoder adopts graph convolutional networks to decode information by reconstructing edges. The nodes of the graph are extracted using a pre-trained CNN network. Experimental results on publicly available clustering datasets demonstrate that GASC outperforms most existing clustering models.