Research On Network Representation Learning For Anomaly Nodes And Dynamic Networks

With the rapid development or Internet tecnnoiogy,Networks,sucn as social networks and citation networks,have been involved in many aspects of people's daily life and works,and contained more and more valuable information.Network Representation Learning generates real-valued,low-dimensional and dense representations for every node in the network,so that they can be applied to various practical tasks such as node classification,link prediction and so on.However,most of the existing network representation learning methods don't consider the screening and refining of network information,but the real network often contains some nodes with anomaly neighbor-hood structure,such as marketing numbers,etc.The existence of these "anomaly nodes"makes it difficult to capture the connections and differences between the ordinary nodes.In addition,most existing network representation learning methods can't use the higher-order neighborhood information of the network effectively.Besides,the networks in the real world are always dynamic,whose nodes and edges will change over time.The methods based on static network representation learning often face the challenges of the structural complexity,variety and evolution timing of dynamic networks.This paper studies the above problems and proposes some solutions.The main research contents are summarized as follows:1.Network Representation Learning Based on Anomaly Node Identification.Considering that the existing Network Representation Learning methods don't take the impact of anomaly nodes in the real network into account,this paper intro-duces an Anomaly Node Identifier into the Network Representation Learning method,to identify and mark the anomaly nodes in the original network based on the network structure information.In order to making full use of the high-order neighborhood information and making the model suitable for the directed network,this paper proposes a Network Attention Autoencoder.It uses multiattention convolution layers to replace the Graph Convolution layer in the traditional Graph Autoencoder,and aggregates the neighbor information based on the probability transition matrix of different orders,to capture the high-order structure of the network.2.Dynamic Network Representation Learning based on Recurrent Neural Network.In order to making full use of the dynamic information of the network,as well as maintaining the stability of the dynamic network embedding,this paper proposes a method to capture the evolution information of the dynamic network using Recurrent Neural Network.Each recurrent cell is a Network Attention Autoencoder with a Gated Recurrent Unit.The proposed model uses the snapshot sequence of the dynamic network as the input,and sets and maintains a state attribute of the nodes to deal with the changing nodes over time.Besides,considering the ability on processing time sequence information of Gated Recurrent Unit,our method uses it to mine the dynamic network evolution information and obtain the stable node embeddings.The proposed methods are applied to node classification,link prediction and community discovery tasks respectively.Experiments are carried out on multiple datasets.The experimental results show that the proposed methods achieve better results on all experimental datasets compared with the benchmark methods.