Research On Dynamic Community Evolution Detection Method Based On Attribute Node Network Representation Learning

In recent years,research in computer-related fields has developed rapidly.As a hot topic in the research community,complex networks have attracted many scholars to conduct a large number of studies on related issues.Among them,the problem of dynamic community detection is of great theoretical value and practical importance for better analysis and prediction of the behavior of individuals in a network,so there has been a high level of research and discussion on this problem in recent years.The problem of community detection refers to the process of partitioning and clustering individuals in a network by similarity or fixed criteria.However,the existing results mainly focus on static networks,which are directly clustered by graph partitioning,hierarchical clustering,heuristics,and other methods.In the real world,where the nodes and edges in the network change at all times,related studies in static networks can no longer meet the needs of existing research,and thus the field of dynamic community detection research has emerged.Currently,the study of dynamic community detection is almost entirely based on the improvement of static research methods.By slicing the network at fixed time or fixed data length,community partitioning results are studied and analyzed in continuous-time slices.However,most of the previous methods for dynamic community detection ignore the node relationships between consecutive time slices and the attribute information contained in the nodes,which leads to low efficiency and division accuracy of the algorithm when studying multiple time slices.Therefore,it is of practical interest to integrate the external information of nodes in the process of dynamic community detection to improve their segmentation accuracy and smooth transitions between each time slice.To address this issue,in this paper,we introduce the idea of network representation learning and topological potential in the development of research methods.The combination of the two can not only efficiently learn node feature representations with attribute information,but also compensate for the shortcomings of repeated training between consecutive time slices.We design and implement a dynamic community evolution method based on attribute node representation learning.?(1)In dynamic network representation learning,in order to make the trained node vector contain more abundant information.First,the idea of Node2 Vec is used to perform random walks in time slices to collect sequences of sampled nodes with network structure information.Secondly,the attribute information of the nodes is obtained by the One-Hot idea.Finally,the Skip-Gram model is used to maximize the likelihood probability of the sequence of nodes formed by the above random walk and the One-Hot strategy.Combining the above two random walk strategies,we propose a random walk based representation learning method that takes into account node attributes.The proposed method is validated in real-world networks using evaluation metrics.The results show an average improvement of 31% in time efficiency compared to the comparison algorithm.?(2)Considering that the traditional incremental method only focuses on the incremental node,but in fact the neighbor nodes of the incremental node will also be affected and the community affiliation will change.To carry out the next related work on dynamic community detection more effectively,it is necessary to introduce topological potential field theory and propose an incremental community detection method.First,the optimal influence range value is judged and determined by the variation of the entropy of the topological potential.Second,on two adjacent time slices,the new or reduced nodes or the new or reduced edges in the time slice at time t + 1 and the nodes and edges within their influence range are traversed and re-trained to form a new node vector.Then,the similarity value between nodes is computed from the cosine similarity of the node vector and the edge weights are updated.Finally,the network is partitioned into communities using the modularity gain idea,and the partitioning results are obtained.The time efficiency of the proposed method is improved by an average of15.2 percent compared to the traditional dynamic community partitioning method,which demonstrates the feasibility of introducing incremental and network representation learning methods.