Research On Knowledge Discovery In Complex System

With the development of deep learning and big data in recent years,complex system disciplines with decades of research history have shown new research opportunities.The study of complex systems is widely used in various fields such as biology,chemistry,computers,communications,and social sciences.Since the characteristics of complex systems exist widely in nature,human society,and artificial systems,the pattern discovery and prediction problems in these actual systems can be summarized as knowledge discovery problems in complex systems,including community discovery,link prediction,node prediction,graph prediction and etc.Due to the universality of its tasks,knowledge discovery in complex systems has wide application scenarios.Based on the above background,this thesis combines the structural characteristics of complex systems with knowledge discovery tasks,and aims to conduct a comprehensive study on knowledge discovery in complex systems.This thesis first gives the general model of complex systems and the definition of knowledge discovery tasks.And based on the definition,this thesis conducts three aspects of research,pattern detection based on unsupervised learning,global classification based on supervised learning and local pattern prediction.A general neural network architecture of knowledge discovery in complex systems is proposed at the end of the thesis.1.Unsupervised detection of complex network patternsComplex networks often exhibit the characteristics of multi-layer heterogeneous networks,and there are interactions between multiple layers.Based on this,this thesis conducts pattern discovery research on mobile communication networks and proposes a spatial pattern discovery algorithm based on spectral clustering.First,an individual characteristic extraction method is proposed.To build a complex network model for mobile communication systems,this thesis uses the individual characteristics of base station to construct the network.This network retains mobile communication characteristics from various data.The spatial correlation contained in the system lays the foundation for the subsequent reseach of complex network knowledge discovery.Second,this thesis studies the complexity characteristics of mobile communication systems.The mobile communication system is a two-layer coupled complex system,which contains mobile user networks and base station networks.According to the data analysis of the current network,the mobile communication system has obvious nonlinear characteristics of complex networks.Finally,a spatial pattern detection algorithm based on spectral clustering is applied to this network and experiments are conducted.Experimental verification shows that this algorithm can adaptively adjust the number of clusters to find base station communities with different characteristics at different spatial scales.This research result has certain theoretical guiding significance for the telecom operators in the network deployment,operation and maintenance,and service promotion of the mobile services.2.The classification of complex network global patternTo solve the pattern prediction problem of complex networks,this thesis proposes a classification algorithm for complex networks based on representation learning.Researches in the field of traditional complex networks mainly focus on the statistical or dynamic characteristics of a single complex network,and few studies on the global classification of different complex networks.The complex network features are usually in a high-dimensional non-euclidean space,which makes the traditional machine learning classification algorithm unable to be directly applied to this problem.In order to predict the overall mode of complex networks,this thesis proposes a complex network classification method based on network representation learning.First,the network embedding algorithm converts the node features into vectors and embeds them into the high-dimensional vector space.Then,by using PCA,the feature reduction is performed,and the high-dimensional space is reduced to the two-dimensional Euclidean space to form images.The images are trained by convolutional neural networks to achieve the purpose of complex network classification.At the mean time,an entropy-based PCA optimization algorithm is proposed for dimensionality reduction after the last layer of convolutional features that can be applied to traditional classifiers.The research can be applied to community diagnosis of complex networks and prediction of complex networks.3.Local pattern prediction in complex networks and GNN architectureFor the prediction of local patterns in complex systems,this thesis proposes a graph pooling algorithm based on feature collaboration and a graph neural network architecture for spatial pattern detection.First,the pooling component of feature collaboration is used to aggregate the local patterns in the network into nodes.Pooling operations extract more abstract features in the network.Second,the algorithm combines the weighted walk based on multiple edges to complete the node prediction task.Experiments show that the weighted walk on multiple edges can effectively reduce the amount of training parameters in the classifier while integrating more features.The thesis applies the algorithm to three citation data sets for experimental verification.The results show that the prediction accuracy of the algorithm is improved compared with other graph neural network algorithms.At the end of this article,we discuss the architecture of deep neural network and give the general architecture of spatial pattern detection network,including the feature extraction component responsible for node representation,the pattern recognition component responsible for community discovery,and the pattern alignment component responsible for network embeding and the classifier responsible for prediction.This architecture can implement different knowledge discovery tasks by using different components.The research results of this thesis can be applied to the construction of a general system of complex system knowledge discovery to achieve rapid engineering applications.