Research On Representation Learning Of Multi-layer Networks

In order to accurately describe the complex and diverse relationships between objects in real systems,multi-layer networks are often used,where each layer represents a type of relationship between nodes.Network representation learning can independently learn a mapping function,embedding the nodes into a low-dimensional vector space while retaining the rich information of the original network,which can be used for downstream tasks such as link prediction and community detection.Traditional network representation learning methods are mainly proposed for single-layer networks,and the research on multi-layer network representation learning is still insufficient.The main challenges include:(1)how to promote learning by inter-layer collaboration;(2)how to enhance the representation of sparse layers by handling inter-layer imbalance.To address the above problems,this paper systematically carried out the following works.(1)Inspired by the single-layer network representation learning method based on the marginal probability model,we propose a multi-layer network representation learning algorithm based on node context coordination(MECNC).Firstly,in order to keep different layers in the same semantic space,node context is defined and used as a bridge connecting each layer space to learn the node representation collaboratively;Secondly,the variance is used to make the representation of the same node in different layers as close as possible,and further integrate the representations in different layer spaces;Finally,the node representation is divided into out-degree representation and in-degree representation,so that the model can adapt to both directed and undirected graphs.Experimental results show that MECNC outperforms baseline methods in link prediction performance on five publicly available datasets.(2)We divide the multi-layer network into target layer and auxiliary layers and propose a layer-imbalance aware multi-layer network representation learning algorithm(USIMNE)based on auxiliary layers under-sampling strategy.Firstly,the common features of nodes are obtained through network aggregation and random walks;Secondly,the dense auxiliary layer is under-sampled based on the node similarity to retain information related to the target layer and achieve inter-layer balance;Finally,an extended Node2 vec method is used to learn node representations for the multi-layer network.Experimental results show that the proposed method outperforms other single-layer and multi-layer network representation learning methods on the link prediction task in the target layer.