Bipartite Network Representation Learning Algorithm Research And Application

In the real world,relationships generally exist between things,where relationships between different types of things are suitable to modeled with bipartite graphs.They can be used to identify or represent a single thing in immeasurable value.In order for computers to understand and make full use of these relationships,we need to learn a vector representation for each node,making it available to other machine learning models.Therefore,it is necessary to fully consider properties of bipartite graphs to design models suitable for them.This paper first investigates the current status and defects of graph representation learning models,and finds that they are not applicable to bipartite graphs,fail to model the explicit and implicit relationships in bipartite graphs,or have limited capabilities due to linear structures.Later,this paper proposes three bipartite graph embedding models.The first is a non-deep model based on explicit and implicit relationships,which can capture both explicit and implicit information in the graph.The second is a model based on a graph convolutional network,which is a en-decoder structure.Each layer of the encoder performs message update and aggregation operations to capture neighborhood information,so it has non-linear modeling capabilities.The third is a model based on convolutional networks,which receives features or hidden representations of random selected nodes,and can be performed on large-scale graphs.The experimental results show that all of three models have different application scenarios,perform well and meet their design expectations.