【作者】 陈超；

【导师】 陆虎；

【作者基本信息】 江苏大学 ， 计算机技术（专业学位）， 2022， 硕士

【摘要】 随着大数据时代的到来,人们越来越急需一种聚类算法来对大量的无标签数据进行处理,但我们往往面对的样本数据量较大,数据维度较高,采用传统的聚类算法不仅耗时长,而且聚类结果往往也不尽如人意。同时,聚类是一个迭代优化的过程,聚类过程中往往会受到噪声数据的干扰而导致聚类结果的性能较差。本论文针对上述的主要问题,主要从模型结构、损失函数、聚类算法、样本筛选等方面,研究了基于深度自编码器的三种新方法,具体的研究内容如下:第一,本论文提出了一种基于样本再选择的深度嵌入聚类算法(IDCEC)。该算法在深度聚类之后对样本进一步筛选,挑选出部分带有伪标签的可信赖样本传入到卷积神网络中进行训练,进而得到一个更好的分类模型。同时,本论文设计了自己的深度聚类模型结构和对应的损失函数。模型及结构上,采用的是带有池化层和上采样层的卷积自动编码器以及具有软标签分配能力的聚类层。根据深度聚类模型的重构损失、聚类损失以及权重和偏置的正则化约束项提出了新的损失函数,提高了模型的特征提取能力与最终的聚类效果。第二,本论文在第一点工作的基础上,融入了自表达层和三元组的样本筛选机制,提出了一种深度的端到端结构的自监督子空间聚类模型(DSCNSS)。该模型利用自监督的思想,配合前置任务中的子空间聚类模块,使得模型的自表达系数矩阵以及聚类结果得到充分的利用。同时,本论文针对该模型提出了一种新的交替优化策略。不同于传统的联合优化策略,该方式在训练的过程中既可以侧重学习前置任务中的自表达系数矩阵从而辅助下游任务进行自监督学习;也可以侧重学习下游任务中的分类模块,反向优化模型的特征提取部分。第三,本论文基于前两个工作内容,提出了一种基于自监督图卷积自动编码器的深度聚类网络模型(DS-AGC)。该模型凭借自身的图注意力自动编码器和聚类来获取样本的伪标签,然后采用了本论文提出的可信赖样本筛选机制,将置信度较高的样本及其伪标签传入到下游任务的图卷积神经网络中,以供下游模型进行训练,最终得到样本的预测标签。模型实现一个无监督的图聚类方法。该模型不依赖样本标签,可借助数据自身进行自监督学习产生伪标签,提高聚类性能。更多还原

【Abstract】 With the advent of the era of big data,there is also an increasingly urgent need for a clustering algorithm to process a large number of unlabeled datasets.However,the samples faced are often large in data volume and high in data dimensionality,and the use of traditional clustering algorithms is not only time-consuming,but also the clustering results are often unsatisfactory.At the same time,clustering is an iterative optimization process,and the clustering process is often disturbed by noisy data and leads to poor performance of clustering results.In this thesis,three new methods based on deep auto-encoders are studied in terms of model structure,loss function,clustering algorithm,and sample selection to address the main problems mentioned above,as follows:First,this thesis proposes a deep embedding clustering algorithm(IDCEC)based on sample reselection.The algorithm further screens the samples after deep clustering,selects some reliable samples with pseudo-labels to pass into the convolutional neural network for training,and then obtains a better classification model.Meanwhile,this thesis designs its own deep clustering model structure and the corresponding loss function.The model and structure are based on a convolutional autoencoder with pooling and upsampling layers and a clustering layer with soft-label assignment capability.New loss functions are proposed based on the reconstruction loss,clustering loss,and regularization constraint terms for weights and biases of the deep clustering model,which improve the feature extraction capability of the model and the final clustering effect.Compared with the existing deep embedding clustering algorithms,the deep clustering model in this thesis can obtain more accurate clustering results.Second,based on the work in the first point,this thesis incorporates the selfexpression layer and the sample selection mechanism of triples and proposes a deep end-to-end structured self-supervised subspace clustering model(DSCNSS).The model utilizes the idea of self-supervision with the subspace clustering module in the predecessor task so that the self-expression coefficient matrix of the model and the clustering results can be fully utilized.Meanwhile,this thesis proposes a new alternating optimization strategy for the model.Unlike the traditional joint optimization strategy,this approach can focus on learning the self-expression coefficient matrix in the predecessor task to assist the self-supervised learning in the downstream task,and also focus on learning the classification module in the downstream task to optimize the feature extraction part of the model in reverse.Third,based on the first two works,this thesis proposes a deep clustering network model(DS-AGC)based on a self-supervised graph-convolutional autoencoder.The model obtains the pseudo-labels of samples by virtue of its own graph attention autoencoder and clustering and then adopts the reliable sample selection mechanism proposed in this thesis to pass the samples with higher confidence and their pseudolabels into the graph convolutional neural network of the downstream task for training by the downstream model to finally obtain the predicted labels of the samples.The model implements an unsupervised graph clustering method.The model does not rely on sample labels and can improve the clustering performance by generating pseudolabels with the help of self-supervised learning of the data.更多还原