A Novel Compressed Sensing-Based Graph Isomorphic Network for Key Node Recognition and Entity Alignment

2.1 Entity Alignment

The traditional entity alignment methods mainly include the paired entity alignment algorithm and the collective entity alignment algorithm. In the paired entity alignment method, the traditional probabilistic model based alignment method is a matching method based on attribute similarity, which transforms the entity matching problem into a classification problem through attribute similarity score. Zhuang et al. (2016) analyzed that the collective alignment method based on similarity propagation, which comprehensively utilizes the attributes and relations of entity pairs, has better accuracy and expansibility. The method based on similarity propagation can truly realize collective alignment.

In recent years, many scholars have focused on the research of graph neural network. This kind of method generally uses the recursive aggregation method to encode the information of the neighbor nodes around the node so that the structural information of the graph can be used in the subsequent alignment process. Sun (2020) proposed a novel knowledge graph-aligned network, AliNet, which alleviates the non-isomorphic neighborhood structure in an end-to-end manner. Xu (2019) proposed a topic entity graph to represent entities' knowledge graph contextual information, using a graph convolutional neural network to encode the two graphs. Some scholars have also proposed relying on graph similarity for graph matching. For example, Bai (2019) proposed a new attention mechanism to emphasize important nodes. The model has been well promoted on invisible graphs. Li (2019) trained graph neural networks to handle various supervised prediction problems defined on structured data, enabling efficient similarity inference. In addition to the above methods, scholars have used other dimensional information for graph embedding to improve the alignment task. Chen (2018) introduced an embedding-based method for entity description-based semisupervised cross-lingual learning using weakly aligned multilingual knowledge graphs. Yang (2019) proposed a novel approach to learning cross-lingual entity embeddings, which incorporates multifaceted features. K. Yang et al. (2020) proposed the COTSAE model, which combines the structure and attributes information of the entity by jointly training two embedded learning components. Zhu et al. (2021) proposed a new framework based on relation-aware graph attention networks to capture the interactions between entities and relations. Xu et al. (2021) embeds entities, relations, and timestamps of different knowledge graphs into a vector space for entity alignment.