Graph Structure-based Multi-source Transfer Learning

In recent years,with the rapid development of big data and artificial intelligence technology,more and more network related applications such as social network and citation network have been emerged.A unique aspect of these applications is that the data is represented by a network structure,where nodes represent entities and links represent relationships.By aggregating a large amount of graph network knowledge to realize rapid response and reasoning of knowledge.However,the new network doesn't be able to obtain a good classification model by using the machine learning algorithm due to have no available labels or acquire node labels cost high.But often,rich markup data usually exists in many established networks from different but related domains.This study uses the information from the secondary source network to accurately predict the target network nodes' labels.In reality,the main challenge of transfer learning across networks is to identify knowledge that can migrate between networks and useful for target networks.For data with network structure,a graph structure-based multi-source transfer learning is proposed in this study.This method constructs node structure features,combines structure features with node entity features,and uses iterative algorithm to classify nodes in target domain.This study proposes a method to construct the node structure characteristics of data with network structure.In the local neighborhood structure of all nodes,a set of graph structures that are effective for classification is found.The similarity between the local neighborhood structure of nodes and this set of graph structures is calculated as the structural characteristics of this node.For the graph structure involved in this study,a maximum common subgraph solution method is proposed,which can effectively filter the set of candidate subgraphs and calculate the results in a limited time,which meets the design requirements of the algorithm in this paper.A large number of experiments in open datasets and real networks show that the algorithm proposed in this study has better performance than the other baseline approaches,It also shows that the accuracy of node classification can be improved by transferring inductive structure information between networks.