Research On Fusion Learning For Graph Clustering

When dealing with massive amounts of data with diverse sources,various types,complex structures,and enormous scales,how to efficiently extract the internal structure information of data and effectively classify and categorize the samples is a problem that data science experts need to pay attention to.Clustering methods,as an essential paradigm in unsupervised learning,can solve the problem of scarce data labeling and therefore have gained popularity,while graph clustering method has greater flexibility and adaptability when dealing with complex,multi-source,heterogeneous and low-quality data.Therefore,in the process of processing complex and multi-source data,graph clustering methods have become one of the most important data analysis tools.However,the existing graph clustering methods still have many defects and deficiencies,such as weak noise robustness,low-quality initial graph,insufficient acquisition of nonlinear relationships in the data,lack of ability to mine complementary information from multi-source data,and the tendency to accumulate bias in multi-step learning strategies.Therefore,how to construct a data graph with strong adaptability,excellent generalization and high stability is the most critical problem in the process of graph clustering modeling.This dissertation focuses on the application of fusion learning in graph representation learning and aims to propose a more robust and stable graph clustering method to overcome the defects and deficiencies of the existing methods so as to improve the clustering performance of the whole model.The main innovations and contributions of this dissertation are as follows:(1)To address the problems such as weak robustness to noise and poor quality of initial graphs,this dissertation proposes an adaptive dynamic graph clustering method based on bistatic graph fusion.Due to the traditional artificial method of constructing initial graph and the setting of hyperparameter,the quality of data graph is often low,which leads to the instability and uncertainty of the final clustering effect.This dissertation adaptively learns a dynamic graph with a suitable sparsity between the sparsest static graph and a dense static graph by using fusion learning.In this method,bistatic graph fusion terms and rank constraint terms are skillfully combined in a joint optimization framework so that the method can adaptively learn dynamic graphs with definite connected component constraints from multiple initialized static graphs.Meanwhile,during the whole process of model optimization,the dynamic graph keeps the most suitable connected structure for clustering requirements,and avoids the cumulative deviation problem of multi-step learning.(2)To address problems such as insufficient exploration of non-linear relationships in data and inadequate representation capability of graphs,this dissertation proposes a self-representation graph clustering method based on minimum-maximum optimization strategy.Because most of the real data has nonlinear characteristic relationship,the traditional simple static graph may not be able to capture all the internal structures,especially the nonlinear structures between samples.To effectively intergrate and learn these multiple candidate base kernels,this dissertation proposes a game-theoretic minimum-maximum optimization strategy to obtain a consensus kernel that can describe the multiple structure characteristics of data more comprehensively.Then,we learn a more comprehensive self-representation graph of the sample structure in the feature space of the consensus kernel.This method integrates such subtasks as multi-kernel combinatorial optimization,self-representation graph learning,clustering structure division into the unified framework of collaborative optimization,and effectively solves the nonlinear representation problem in traditional graph clustering methods.(3)To address problems such as lack of mining ability for complementary information from multiple sources,this dissertation proposes a multi-source graph fusion clustering method based on contrast feedback optimization.Complex and diverse data generally have problems such as low quality and heterogeneity,which undoubtedly increases the difficulty of seamless fusion of multi-source data and also makes the traditional model more prone to instability and difficult convergence in the learning process.Therefore,this dissertation first converts the heterogeneous features of the original data into high-order homogeneous feature representations by constructing local neighborhood graphs in the preferred kernel space,thereby eliminating differences in data types and configurations to handle the complex and diverse sources of multisource heterogeneous data.Then,the mutual information comparison and feedback strategy is used to iteratively optimize the kernel graph of each information source and its dynamic graph neighbors after fusion learning by constantly using the complementary and consistent information among multi-source data to ultimately improve clustering performance.This dissertation aims to improve the existing graph clustering model,design a more stable and effective graph representation,and apply it to the clustering method based on graph learning.To achieve this goal,we combine the architectural basis of fusion learning with the graph clustering model,gradually introduce multi-kernel learning,gamification learning,feedback learning and other methods so as to constantly update,discover and mine the distribution characteristics of samples,obtain the optimal graph structure,and ultimately complete the clustering application of sample data.