| In recent years,deep learning,especially deep supervised learning,has developed rapidly,but it is constrained by data challenges.The so-called data challenge is that with the advent of the era of big data,it is easy to obtain a large amount of data,but it is time-consuming to obtain the corresponding data labels.Therefore,deep semi-supervised learning for training deep neural networks using both labeled and unlabeled samples has become a hot research topic.The key to solving the data challenge is to make full use of unlabeled data.Recently,self-supervised contrastive learning,as a technique that utilizes the information of the data itself,has shown good performance in the field of computer vision,which indicates that it can effectively avoid the dependence of algorithm performance on a large number of labels.In addition,there are also data challenges in the field of bioinformatics,which mainly lies in how to transform and integrate some prior knowledge information into existing unsupervised learning algorithms.Taking single-cell RNA-sequencing(sc RNA-seq)data analysis as an example,some biological prior knowledge is relatively easy to obtain,but existing single-cell RNA-sequencing data analysis algorithms seldom utilize the prior knowledge.Although some semi-supervised clustering has integrated prior knowledge into the clustering process,the prior knowledge used is still local.In semi-supervised learning algorithms,unlabeled data are mostly utilized by consistency loss or data reconstruction loss.Those methods utilize less unlabeled information.In fact,existing deep semi-supervised algorithms exploit unlabeled data to a lesser extent and have less prior knowledge that can be exploited.In response to the above problems,this paper proposes a deep semi-supervised classification algorithm based on contrastive learning and partial label propagation and a deep semi-supervised clustering algorithm for sc RNA-seq based on self-supervised learning and pairwise constraint propagation.For the deep semi-supervised classification algorithm,this thesis proposes a deep semi-supervised classification algorithm CL＿PLP based on contrastive learning and partial label propagation.CL＿PLP consists of two modules,including a feature extraction module and a partial label propagation module.The feature extraction module adds a projection layer module to the network structure,constructs a symmetric contrastive loss and expand the dataset by combining strong and weak augmentation.The partial label propagation modules interrupt label propagation according to the quality of pseudo-labels to generate high-quality pseudo-labels.This paper compares with the state-of-the-art algorithm on three standard datasets,and the results show that the algorithm improves the utilization of unlabeled data and obtains better classification results.This thesis also applies the algorithm to the COVID19-Xray dataset,which still shows good performance.In addition,this thesis conducts ablation experiments to analyze the impact of different modules of the algorithm on the performance.Finally,we propose a strategy to integrate this algorithm with an inductive learning method,and it is experimentally demonstrated that this strategy can further improve the performance and obtain additional high-quality pseudo-labels.For sc RNA-seq deep semi-supervised clustering,this thesis proposes a new sc RNA-seq deep semi-supervised clustering algorithm sc DSemi C based on self-supervised learning and pairwise constraint propagation.sc DSemi C first pre-trains an autoencoder with two classes of contrastive learning.Then use the pre-trained model to extract data features,construct adjacency matrix and constraint matrix for pairwise constraint propagation.Finally,semi-supervised constrained clustering is performed using the propagated high-quality constraints and the original constraints.This thesis verifies the effectiveness of the algorithm by comparing it with five competing algorithms on seven real sc RNA-seq datasets,and analyzes the sensitivity of the key parameters and the effectiveness of different combination modes of modules. |
