Research On Link Prediction In Complex Networks Based On Graph Representation Learning

In the real world,many networks consisting of entities and their relationships can be referred to as complex networks.Generally,complex networks are network structures composed of many nodes and complex relationships between nodes,such as social networks,transportation networks,drug target interaction networks,and so on.Link prediction is an important research problem in the field of complex networks,and its core task is to predict links in the network that have not been observed or may occur in the future based on the existing network structure and information.The continuous evolution and complexity of complex networks have made link prediction an important and challenging problem in the field of complex network analysis.In this paper,we propose corresponding link prediction models for two typical representatives of complex networks: drug-target interaction networks and multi-layer dynamic social networks.The main contributions of this paper are as follows:(1)A graph representation learning-based mixed expert drug target interaction prediction model(GRMEDTI)is proposed to address drug-target interaction prediction problem in the drug-target interaction network,which includes multiple entity types and relationship types.Existing methods usually adopt network embedding techniques to learn the feature representations of entities,but they overlook the type prior information of the links between entities and tend to isolate the feature representations of a certain entity type while ignoring the correlations generated by interactions with other entities.Therefore,this method captures the relationships between the same and different types of links by using type prior information,which is applied to drug target prediction.Experiments on two drug target datasets show that the proposed GRMEDTI model can effectively distinguish different relationship types between entities,extract corresponding link type prior information,and enhance the embedding representation of different types of entity relationships,which leads to accurate drug target prediction.The prediction accuracy of our model outperforms the current state-of-the-art algorithms.(2)We propose a link prediction method,named Multi-layer Temporal Perception Graph Representation Learning Method(MTPGRM),for multi-layer dynamic social networks.Existing methods focus on analyzing either multi-layer or dynamic networks,but do not consider the correlations between the multi-layer structure and the temporal dynamics of social networks.To address this issue,MTPGRM is proposed to comprehensively consider the multi-layer structural and temporal dynamic correlations of social networks.By using multi-layer temporal perception graph representation learning,MTPGRM learns the graph representation of the multi-layer dynamic social network and obtains the embedding representation of network nodes and links.Then,the true value discovery algorithm is used to integrate the prediction results of unobserved links on the auxiliary layer and target layer,which is further fed back to predict unobserved links on the target layer.Finally,the Kalman filter is used to correct the prediction values.Experimental results on real-world datasets demonstrate that our proposed method outperforms state-ofthe-art methods in link prediction for multi-layer dynamic social networks.