Deep attribute graph clustering based on bisymmetric network information fusion and mutual influence

Deep attribute graph clustering based on bisymmetric network information fusion and mutual influence

연구 분야: Databases

학회: Applied Intelligence

Deep attribute graph clustering has always been a challenging task and an important research topic for real-world data. In recent years, there has been a growing trend in using multi-network information fusion for deep attributed graph clustering. However, existing methods in deep attributed graph clustering have not effectively integrated representations learned from multiple networks and failed to construct a joint loss function that could impact the overall network model, resulting in poor clustering results. To address the aforementioned issues, we proposed AGC-BNIFI, an attribute graph clustering method based on dual symmetric network information fusion and mutual influence. The network of this method consists of a symmetric graph autoencoder and an autoencoder. The two different encoders are combined to improve the attribute learning ability. First, a symmetric graph autoencoder with a symmetric structure is proposed to capture complex linear and adapt to complex graph structure relationships and propagate heterogeneous information of joint embedding and structural features, and can reconstruct the attribute matrix and adjacency matrix; secondly, a layer-by-layer adaptive dynamic fusion module is designed to adaptively fuse the representations learned by each layer of the two encoders, and then learn a better joint representation for clustering tasks; finally, a multi-distribution self-supervision module with soft clustering assignments obtained from different networks that learn from each other and influence each other is proposed, which integrates representation learning and clustering tasks into an end-to-end framework, and jointly optimizes representation learning and clustering tasks by designing a joint loss function. Extensive experimental results on four graph datasets demonstrate the superiority of AGC-BNIFI over state-of-the-art methods. On the Coauthor-Physics dataset, compared to MBN, AGC-BNIFI achieved improvements of 2.6%, 1.1%, 4.3%, and 6.3% in four clustering metrics, respectively.