An Embedding Algorithm Based on Neighborhood Aggregation for Attribute Graphs

Attribute graph embedding algorithms based on graph neural networks aggregate information about the neighborhood of a target node to update the embedding of the target node, i.e. neighborhood aggregation. The shallow network architecture of existing models limits the generalization ability of the models on large datasets, and the number of nodes increases exponentially with the number of graph neural network layers, resulting in gradient explosion and the need for faster model training methods. In addition, the multiple iteration process of neighborhood aggregation changes the original attribute information of the nodes, especially when the attributes of the target node and the attribute information of the neighborhood nodes are extremely different, which makes the target node attributes deviate significantly from the original attributes. Excessive neighborhood aggregation mixes different clustered nodes, causing all nodes to converge to the same embedding, resulting in excessive smoothing. In order to solve the above problems while considering the supervised learning of the model and the utility of neighborhood aggregation, this paper proposes a subgraph sampling method based on graph sampling for attribute graph data pre-processing, and an attribute graph embedding algorithm for improving neighborhood aggregation using an adaptive learning strategy.