Research And Application On Community Detection In Bank Customer Transaction Network

Commercial banks have accumulated a large amount of customer transaction data,which contains many potential patterns.Fully mining and using these patterns in customer transaction data can improve the management level and business competitiveness of commercial banks to a higher degree.Complex networks are an effective means of modeling complex systems such as bank customer transaction system.Simultaneously,community structure that is prevalent in complex network provides a feasible starting point for complex network study.We can make full use of community structure to study the topological structure and the dynamic behavior of complex networks,so as to effectively mine the potential patterns of complex systems.The method based on modularity optimization is one of the mainstream methods to detect community in complex networks.Although Louvain Method is an excellent algorithm to detect community structure in undirected network,it still has some defects.First,the algorithm need to calculate the modularity-gain after nodes leave their community in undirected network,however,there is no formula to calculate this gain directly in current study.Secondly,compared with other similar algorithms,it has a higher demand for storage space.Compared to the research on community detection in undirected network,the research about community detection in directed networks is less involved.At present,the feasible method is to make the modularity of directed network as the objective function,then using the idea of modularity optimization algorithm of undirected network to design the corresponding modularity optimization algorithm for directed network.However,most of modularity optimization algorithms for directed network have high time complexity,and the quality of community structure they detected is also unsatisfactory.We constructed five customer transaction networks based on the real customer transaction data of a commercial bank.We also proposed improvements to Louvain Method and designed an algorithm to detect community structure in directed network.Finally,the community structures of five networks were detected and the similarities and differences among them were demonstrated.The main work of this thesis is as follows:(1)We proposed an improved Louvain Method based on a verity of strategies.First,the formula to calculate the modularity-gain after nodes leave their community was deduced,and it was used in Louvain Method so that there was another case that each node belonging to the community in the algorithm.Second,the improved strategy based on separating isolated nodes was put forward to overcome the shortcoming of Louvain Method that has a high demand for storage space.Finally,the refinement strategy was introduced in Louvain Method to further improve the accuracy of the community structure.Relevant experiments based on real networks show that our improved strategies can not only reduce the demand for storage space of Louvain Method,but also make the algorithm run faster and the results of community detection more accurate.(2)We proposed an algorithm to detect community structure in directed network based on Louvain Method.First,the definition for modularity based on the community connection matrix of directed network was proposed.Second,the formula to calculate modularity-gain after nodes merge and the formula to calculate modularity-gain after nodes leave their community were derived.Finally,based on the idea of Louvain Method,an algorithm to detect community in directed networks was proposed.Relevant experiments based on real networks show that the algorithm has obvious advantages both in runtime and accuracy of community discovery results.(3)Community structures of five customer transaction networks were detected by algorithms proposed in(1)and(2).Then,a new modularity was proposed to measure the quality of community structures of customer transaction networks.Finally,the community structures of five customer transaction networks and their similarities and differences were demonstrated through analysis and visualization.