Representation And Community Detection Of Attributed Networks Based On Evolutionary Computation And Deep Learning

The large-scale biological systems,social interacting species,and the Internet in our real world are only a few examples of systems which can be modelled as complex networks whose nodes represent the units and links stand for the interactions between them.Complex network theory and related methodologies provide a promising path for people to reveal and understand the intrinsic operating mechanisms of systems,the dynamics of the member distribution and to predict potential evolving tendency.Attributed networks further integrate the attribute information of nodes with topological structure to better model real-world scenarios and thus significantly improve the performance of semantic mining.The research on attributed network has received a lot of attention from scholars in recent years.As one of the most important part in attributed network analysis domain,community detection helps people understand the underlying structures and patterns of networks,which in turn guides practical applications such as infectious disease prevention and traffic planning.In order to better cope with practical problems,the structural connections and attribute information should be considered synchronously for a reasonable representation of the network.Unfortunately,the traditional representation methods only focus on network topology or node similarity.In view of such situation,this thesis mainly focuses on the study of attributed networks.A series of community detection and network embedding algorithms are proposed based on evolutionary computation and deep learning.The main contributions can be summarized as follows:1.The existing study has shown that the very existence of overlapping communities is one of the most important characteristics of various complex networks,while the majority of existing community detection methods are designed for detecting separated communities in attributed networks.Therefore,a multi-objective evolutionary algorithm based on similarity attribute for overlapping community detection in attributed networks is proposed to deal with above challenges.A modified extended modularity dealing with both directed and undirected networks is well designed.Besides,a novel encoding and decoding strategy is proposed to realize the goal of representing overlapping communities efficiently.The experiments on both synthetic and real-world networks demonstrate that our method can effectively find overlapping community structures with practical significance in both directed and undirected attributed networks.Meanwhile,the number of communities can be automatically determined during the optimization process.2.For the community detection problem,existing network embedding-based methods usually use unsupervised clustering algorithms to cluster the nodes in the reduced dimensional space.However,it is a great challenge to tune the hyperparameters to achieve the optimal performance.In addition,it is quite challenging to make full use of both structural and attribute information.To this end,we combine the network embedding with multi-objective evolutionary algorithms for both separated and overlapping community detection problems in attributed networks.Two objective functions concerning community structure and attribute similarity are carefully designed.Moreover,a heuristic initialization method is proposed to get a relatively good initial population and a novel encoding and decoding strategy is designed to efficiently represent the overlapping communities and corresponding embedded representation.The experiments on both single and multiple attributed real-world networks are conducted to illustrate the superior performance of proposed method.3.The existing works on community detection of attributed networks based on evolutionary algorithms ignore the utilization of node attribute information in the evolutionary process,and the methods tend to create densely-connected but non-homogeneous clusters.To address this issue,a structure-attribute coevolutionary algorithm based on continuous encoding and local merging is proposed.Moreover,we introduce different types of motifs to preserve more semantic information about community structure.Experiments on multiple network datasets show that the synchronous optimization of both attributes and topology significantly improves the performance of the algorithm compared to existing methods,and the obtained results are more practical and meaningful.4.Most existing graph representation learning methods are based on local structure and attribute information,but the high-order features of nodes and community structure are largely ignored.In addition,traditional matrix decomposition methods as well as some heuristic methods require huge computational cost.To address the above challenges,this thesis proposes an efficient deep learning-based representation learning method that considers both structural and attribute information of attributed networks.For single-layer attributed networks,a novel adaptive balance mechanism is designed to integrate these two types of information so that the weights between the structural and attribute information can be determined automatically during the optimization process.Community structure is proposed as a higher-order node similarity metric to deal with the data sparsity issue.For multi-layer networks,a multi-order similarity concept is proposed to fully utilize the interlayer information to obtain better node representation results.The experimental results show that the method has excellent performance on multi-label classification,node clustering and visualization tasks.Parameter sensitivity studies and ablation experiments further confirm the effectiveness of the proposed algorithm.5.The large-scale equipment systems in the modern military battlefield share the distinguished features including complicated interactions between entities and rich attributed information.As a result,it is a great challenge to choose appropriate equipment combination according to different military tasks.To deal with this issue,a multi-objective evolutionary optimization method for equipment combination selection based on attributed network modelling is proposed.A parameterized model for equipment systems based on attributed network theory is designed to combine attribute information from both nodes and edges.On this basis,the problem of equipment combination selection is transformed into a community detection task.Two objective functions in terms of underlying structure of system and contribution rate of task are specially designed.The optimal solution selection strategy based on knee point is investigated.The effectiveness of our proposed method has been validated through experiments.