Research On Deep Clustering Algorithm Based On Contrastive Learning

Clustering is one of the important algorithms in unsupervised learning,which mainly aims at unlabeled datasets and groups data according to some specific standard.The deep clustering algorithm greatly improves the performance of the algorithm by combining the neural network with the clustering algorithm and optimizing the feature space and clustering results at the same time.The deep clustering algorithm mainly solves the problem of how to learn the differentiated representation to produce better results.Contrastive learning can learn higher-dimensional and more essential representative features,which has received widespread attention in representation learning.Due to the good performance of contrastive learning,some algorithms jointly optimize contrastive learning and deep clustering.These methods use the basic framework of contrastive learning to learn the representation and cluster based on this representation.However,contrastive learning emphasizes the importance of data enhancement,distinguishes each instance by making different enhancements of the data as consistent as possible in the feature space,and the clustering goal is to group instances.When combining contrastive learning with deep clustering,if the basic framework of contrastive learning is followed directly,the clustering target will be ignored,and the learned representation may not the best representation of the cluster,which limits the clustering performance.Therefore,the natural groups formed by the similarities between different instances need to be considered when combining.In view of the above problems,this paper proposes two algorithms for different data relations in clustering,and combines contrastive learning with deep clustering to learn clustering friendly representation,so as to improve the performance of deep clustering.The research contents main include the following two aspects:(1)An instance-level contrastive clustering algorithm based on self-paced learning.The algorithm introduces the idea of self-paced learning to cluster in a simple to difficult way.The algorithm consists of two phases.In the first phase,by considering the relationship between data and clusters,the easily distinguished data is initially clustered,and the easily distinguished data in this potential space is distributed near the corresponding cluster center to obtain pairwise similarity relationship.In the second stage,contrastive learning is used for training,so that the indistinguishable samples are gradually easy to distinguish,making the samples within each cluster more compact and the samples between different clusters are far away.Positive and negative examples in contrastive learning are constructed from pairwise similarity results obtained in the first stage,then these examples to conduct instance-level contrastive learning.The experimental results of algorithm on common datasets are improved in different degrees compared with the results of other algorithms,indicating that algorithm can obtain better clustering effect.(2)An instance-cluster level contrastive clustering algorithm based on graph structure.This algorithm achieves contrastive clustering by considering the relationship between data and data,and between clusters,and upgrades the common instance level to cluster level.Graph structure is used to capture the potential relationship between data.The basic idea of deep subspace clustering is used.The self-expression layer is added to auto-encoder to get the graph structure,from which the neighborhood information of the sample is reflected.Construct positive and negative examples in contrastive clustering by the graph relationship,and use the columns of the characteristic matrix as cluster predictions for data,and perform graph-based cluster level contrastive learning in column direction.Potential class information is fused at both the instance level and the cluster level,feature learning and cluster assignment are trained from two aspects simultaneously.The algorithm is experimented on three datasets and compared with several advanced clustering algorithms to verify the feasibility and validity of the algorithm.This paper proposes two deep clustering algorithms,which combine contrastive learning with deep clustering better by considering the relationship between data from different perspectives,and learn clustering friendly representation to improve the clustering performance.It has important theoretical significance and use value.