Research On Context-based Network Representation Learning Algorithm

With the rapid development of the information society,the data structure of the network exists in more and more realistic situations,and is widely used in computer and related fields.The analysis of these networks has high academic value and practical application value.It is worth noting that effective network analysis generally depends on how the network is represented.The traditional method of representing the network is usually to use high-dimensional sparse vectors,but nowadays,the number of edges and nodes in a complex network may reach billions.Therefore,using traditional network analysis methods to perform calculations and inferences on the entire network faces many difficulties.Network representation learning aims to map the nodes in the network into low-dimensional dense vector representations.Network representation learning alleviates the difficulty of solving problems caused by the high-dimensional sparsity of the traditional method node representation vector,the good performance demonstrated in downstream tasks has attracted the attention of more and more researchers.However,the existing network representation learning methods cannot handle the context information of nodes well,making the results of network representation learning not very satisfactory.Therefore,this dissertation conducts related research on network representation learning algorithms at home and abroad.According to the two cases of using only structure information and fusing structure information and text information,the context-based network representation learning algorithm is deeply studied.The main research content of this article can be summarized as the following three points:This article first introduces the development history and research status of network representation learning,expounds the concept of network representation learning and explains related evaluation indicators,and introduces a series of network representation learning algorithms based on structural information and a series of network representation learning algorithms that merge text information and structural information.After that,the article summarizes the shortcomings of existing algorithms in the process of obtaining node representation vectors,leads to the two algorithms based on node context proposed in this dissertation,and explains the improvements compared to traditional algorithms.This dissertation proposes a Path-based Mutual Attention Network Representation Learning(PMA-NRL)Algorithm.Random walk is used to capture the structural context information of nodes in the graph,and then obtain the mutual attention information of different neighbor node contexts through the attention mechanism,which effectually distinguishes the roles that nodes play when interacting with different neighbor nodes,thereby obtaining more accurate network representation vectors.The algorithm mines the context of the network structure,effectively solves the problem of how to mine deeper information of the network structure when the text acquisition is difficult or insufficient through the random walk algorithm and the attention mechanism,and generates high-quality node representation vectors without using attribute information.The algorithm has achieved good results on three datasets,which shows the advantage of PMA-NRL.This dissertation proposes a Text-based Context Expansion Network Representation Learning(TCE-NRL)Algorithm.Aiming at the traditional network representation learning algorithm only considering the text information of local neighbors or node information reachable after several steps,we propose TCE-NRL to expand the entire graph information into the node context.In the process,node information that is not structurally close to the node itself but similar in text semantics is aggregated.And through encoding and decoding,attributes guide the integration of global network information into the node context,and then aggregate information based on the local structure of the node,seamlessly integrating network information into the node's representation vector.Compared with the traditional network representation learning algorithm,this method obtains a higher quality node vector representation.