Research On Graph Clustering Methods Based On Similarity Learning

Clustering is an important task in computer vision and machine learning.It is widely applied in various fields,such as image segmentation,image classification,scene analysis,motion modeling,and medical image analysis.The clustering problem has been widely investigated by many scholars and a series of excellent approaches have been proposed in the past decades.The graph theory based graph clustering method shows excellent performance.It first models the samples as the weighted undirected graph based on the similarity measure between samples.Then,clustering of samples is achieved by partitioning the graph.Although existing research work on graph clustering technology has made many achievements,there still exist several problems that should be solved urgently.For example,the traditional graph clustering methods generally use a certain function to construct the sample similarity graph,which does not contain any potential clustering structure information and is fixed throughout the learning process.Since the performance of graph clustering method is heavily dependent on the quality of the graph,the traditional approaches of constructing similarity graph will inevitably affect the clustering effect.In addition,most graph clustering methods adopt a two-step learning strategy: the construction of the sample similarity graph and the clustering process of samples are divided into two independent processes.Under such circumstance,there is no interaction between graph generation and clustering process of samples,and it is difficult to obtain the optimal clustering results.To solve these challenges,this paper proposes the similarity learning based graph clustering method.Our work mainly includes two folds:(1)A robust structured graph clustering model is presented.In this model,a unified learning framework is designed to simultaneously cluster samples and learn a robust structured graph.In this method,a new embedded representation is first learned for the samples form the original data to eliminate the interference of noises and outliers from the original data.A sample similarity graph is then learned by adaptive neighbors assignment based on the new feature representation.By imposing a rank constraint to the Laplacian matrix of the graph,the number ofconnected components in the graph is exactly equal to the number of clusters,and the optimal clustering structure is obtained.Finally,the discrete clustering labels of the samples can be obtained directly by partitioning the learned structured graph without any label discretization strategy.To optimize the objective function,this paper first transforms the objective formula into an equivalent one that can be tackled more easily.Then,an iterative optimization method based on augmented Lagrangian multiplier is used to solve the problem.Experiments on multiple datasets show that the proposed method has better performance than the existing clustering methods.(2)A multi-view clustering method based on the dynamic graph feature learning is proposed.In this method,the dynamic graph learning and the feature extraction are simultaneously performed,and the extracted features are directly used for clustering.Among them,the dynamic graph learning process can adaptively capture the intrinsic multiple view-specific relations of samples,and the feature extraction part can preserve the manifold relations of the samples in the original space to the low-dimensional space by learning a projection matrix.This paper proposes an efficient optimization algorithm to obtain the optimal solution of the objective function,and proves the convergence of the algorithm.Compared with state-of-the-art multi-view clustering methods based on feature learning,the proposed algorithm achieves better performance on multiple public datasets.