1. Tree structure perception graph representation learning based on hierarchical aggregation and relational metric learning
Tree Structure-Aware Graph Representation Learning via Integrated Hierarchical Aggregation and Relational Metric Learning
Ziyue Qiao, Pengyang Wang, Yanjie Fu, Yi Du, Pengfei Wang, Yuanchun Zhou
Although graph neural network (GNN) shows advantages in learning the node representation of homogeneous graphs, it is still a challenging problem to use GNN on heterogeneous graphs. The main reason is that GNN learns node representation by aggregating neighbor information regardless of node type. Some work has been proposed to mitigate this problem by sampling neighbors with different categories using relational or meta-paths and then using attentional mechanisms to understand the importance of the different categories. However, one limitation is that the representations learned for different types of nodes should have different feature Spaces, and all of the above work still projects node representations into a feature space. In addition, after studying a large number of heterogeneous graphs, we discover the fact that multiple nodes of the same type are always connected to nodes of another type, which reveals the many-to-one architecture (also known as hierarchical tree structure). But none of the above work preserves the tree structure because the multi-hop relationship from neighbor to target node is eliminated by the aggregation process. Therefore, in order to overcome the limitations of literature, we propose t-GNN, which is a tree structure perception graph neural network model for graph representation learning of tree structure representation. Specifically, the proposed t-gnn consists of two modules :(1) integrated hierarchical aggregation module and (2) relational metric learning module. The integrated hierarchical aggregation module aims to preserve the tree structure by integrating the tree hierarchy and sequential neighborhood information into the node representation by combining GNN with GRU. The relational measures learning module aims to preserve heterogeneity by embedding nodes of each type into type-specific Spaces with different distributions based on similarity measures.
Machine learning: Deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning, deep learning851320808To join the wechat group, please scan the code:Copy the code