| As we all known,the 21 st century is the century of information.With the increasing maturity of instant messaging devices and the increasing popularity of Internet applications,we are entering the era of Information with the explosive growth of data.When dealing with massive amounts of data with great value,how to organize the data and dig out valuable information effectively is particularly important.The network provides a natural and powerful way to represent relational information among data objects from complex real-world systems.Due to the flexibility of using networks to model data,many real-world applications mine information by analyzing networks.One way to analyze networks is to identify the groups of nodes that share highly similar properties or functions,such as proteins with similar functions in biological networks,users with close relationships in social networks,papers in the same scientific fields in citation networks.The research task of discovering such groups of nodes is called the community detection problem.Due to the complex and diverse of real-world applications,it is common that one node may belong to several communities.Therefore,research on community detection algorithms based on overlapping communities has great application value and practical significance.Whats more,community detection in previous models focuses on a single and fixed network.Even for dynamic networks with the same sizes and attributes,retraining is required to achieve the results.The rapid change and explosive growth of information make real-world applications have great expectations for inductive community detection models that can quickly obtain results.To address this problem,we propose LAA,an inductive community detection algorithm based on label aggregation.LAA assigns a membership label to each node,and the label indicates the membership of the node belonging to the community.Each node propagates the label to its neighbors.At the same time,Each node receives the label propagated by the local neighborhood according to the similarity of the neighboring nodes.Through training a series of aggregation functions,the received label and its own label are aggregated as the final label of the node.The aggregation function explains the influence of the membership information of neighboring nodes on which communities the target node will be divided into,i.e.,the relationship between the generation of communities and the network topology.We are based on the assumption that no matter how the structure changes,such a relationship will not change.In addition,on networks of the same size and attributes,such relationships are supposed to be similar.Therefore,through the trained aggregation function,the labels corresponding to the untrained nodes in the networks can be obtained,and the communities corresponding to the nodes in the network are determined.Finally,we verify the performance of the LAA model through a series of experiments.First,we verified the performance of the LAA model based on supervised learning and unsupervised learning.Then,we use different artificial data sets to simulate changes in the network structure and verify the generalization ability of the LAA model.The experimental results show that the LAA model has superior community division ability and generalization ability. |
