Graph Representation Learning Clustering Based On Graph Adaptive Information And Graph Regularization

With the rapid development of the Internet industry,a large number of unlabeled data with rich features and complex structures have been generated in the daily work.Due to the high cost of artificial marking data,clustering analysis,as a typical unsupervised learning method,mines useful information only based on the degree of association between data,which has attracted the attention of many scholars.Among them,the graph representation learning based clustering method is one of the mainstream research directions in the field.Although different scholars have proposed many graph representation learning based clustering methods in recent decades,there are still some shortcomings that learned representation graph does not have a connected graph structure suitable for clustering,or it is less robust to noise and outliers.Aiming to overcome the above shortcomings and improve the clustering quality,based on the global and local structural information of the data,this dissertation proposes two more robust graph representation learning clustering algorithms by introducing different constraints.The main research results of this dissertation are as follows:(1)To obtain a representation graph with exactly k connected components and make it have a more suitable spatial structure for clustering,we propose a graph representation learning clustering algorithm with adaptive neighbor and graph regularization,named RLANGR.By local distance information and rank constraints,the algorithm can get a representation graph with both local and global optimal structures.Firstly,based on the idea that two samples with closer distances are more likely to come from the same cluster,the algorithm learns a representation graph with the local optimal structure.Then,by introducing a noise constraint term into the algorithm model to restrict the noise and outliers mixed on the dataset,the algorithm improves its robustness to noise and outliers.Finally,it is ensured that the learned representation graph has a globally connected structure suitable for clustering by introducing a rank constraint term.To verify its effectiveness,RLANGR and 11 benchmarks are fully compared on 4 kinds of image and 3 kinds of non-image datasets.In addition,the time complexity of the algorithm is analyzed,and the value of suggestion of parameters involved in the algorithm model is given.The results show that on 7 kinds of datasets,compared with the benchmarks,RLANGR can achieve relatively better clustering results.Thus,RLANGR has great generalization capability on various kinds of datasets.(2)Data is often polluted by noise and outliers mixed on the dataset,making it difficult to discover the real distribution structure.To address this problem,we propose a graph representation learning clustering algorithm with adaptive weighted noise constraints,named RLAWNC.Firstly,based on the fact that noise and outliers have a large reconstruction error in the data joint representation,the algorithm introduces an adaptive weight matrix to constrain noise and outliers,weaken the representation contribution of features polluted by noise and outliers,and reinforce the role of the highly discriminative features in the joint representation.Thus the noise robustness of the algorithm is improved.Then,the algorithm introduces a local distance regularization term to improve the representation coefficients of potential similar data in the joint representation.Finally,the algorithm introduces a local distance metric term to learn the representation graph,which has the local optimal structure while being robust to global structure noise.To verify its effectiveness,RLAWNC and 10 benchmarks are fully compared on 5 kinds of image and 3 kinds of non-image datasets,Besides,the time complexity of the algorithm is analyzed,and the value of suggestion of parameters involved in the algorithm model is given.The results show that on 8 kinds of datasets,compared with the benchmarks,RLAWNC is a clustering algorithm with great generalization capability on different kinds of datasets.