Dual Fusion AutoEncoder for Graph Clustering

Yuanchi Ma, Hui He, Kaize Shi, Zhongxiang Lei, Jinyan Liu, Peng Zhang, Zhendong Niu

Keywords: Graph Clustering, Graph Autoencoder, graph embeddings

Abstract:

Clustering of graphs has been an attractive topic in recent years.
Recent research has focused on deep fusion graph clustering methods, i.e., fusing two different network structures to enhance the performance of clustering by capturing both graph structure information and node feature information. However, this approach is constrained by the representativeness of the chosen neural network and the choice of the fusion mechanism leads to an unpredictable degree of discretization of the learned graph embeddings. It thus becomes crucial to obtain more compact graph embeddings compatible with the clustering task. In this paper, we propose a new end-to-end fusion, dual fusion autoencoder for graph clustering (DFAC) for deep fusion networks. Our model makes full use of the topology and feature information of the graph and is trained simultaneously by multiple components to obtain better graph embedding. Benefiting from our design of a new dual fusion mechanism, this captures cross-modal good embeddings containing node topology and node feature information. Such a design makes it learn relaxed k-means and performs self-supervised training to improve the quality of graph embeddings while reconstructing the graph structure. By optimizing the training process that is in a unified framework, multiple components are mutually beneficial. Experimental results on six publicly available datasets demonstrate the superiority of the proposed method.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3421