A time sequence coding based node-structure feature model oriented to node classification. (1st August 2023)

• Record Type: Journal Article
• Authors: Yu, Ruowang; Xin, Yu; Dong, Yihong; Qian, Jiangbo
• Is Part Of: Expert systems with applications. Volume 223(2023)
• Publication Date: 2023-08-01
• Subjects: Graph data mining -- Node classification -- Graph neural networks (GNNs) -- Environmental context -- Time sequence models (TSMs); Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2023.119872
• Languages: English
• ISSNs: 0957-4174

Abstract

Abstract: Graph data mining is an important method for managing complex systems in artificial intelligence (AI). As an important branch of graph data mining, node classification is widely applied in paper classification in citation networks, user classification in social networks, etc. At present, many graph neural networks (GNNs) have been proposed to realize node classification by obtaining node embeddings encoded by aggregating neighborhood features. Most GNNs use one-hot encoding as the initial node feature when there is no node attribute in the graph. However, one-hot encoding is a global encoding technique that cannot reflect the environmental context of specific nodes. To this end, we proposed a graph structure sequential coding (GSSC) model, including a learnable node-structure feature X, sampler, encoder and classifier, to obtain the structural embeddings which can better reflect the topological structure of non-attribute graphs. For non-attribute graphs, we utilized the learnable feature X as the initial node-structure feature, which could be simultaneously trained with the GSSC model. In addition, we used a decoupling scheme to separate the sampling and encoding processes of GSSC. Therefore, we could use different samplers for sparse and dense graphs. In addition, different sampling orders (i.e., temporality) in node sequences could reflect different semantic associations between the nodes. Therefore, we innovatively employed time sequence models (TSMs) as encoders to encode node sequences with temporality. Based on these TSMs, semantic information of the head node in the node sequence can be captured. The experimental results on five datasets confirmed that our GSSC model performs better than other representative methods in node classification for graphs without node attributes. Highlights: Use learnable X as initial node-structure feature on graph without node attribute. Use decoupled Sampler to sample graphs with different densities. Use TSMs to capture the semantic information of head node in the sequence.