Research On Deep Semi-supervised Clustering Algorithm

In machine learning,clustering is an important learning algorithm.Different from the classification algorithm,the clustering algorithm divides each sample point into different families according to their similarity degree when the provided data has no label.On the contrary,the similarity of data sample points of different clusters is very low.In the process of clustering division,we can't know the sign of its division,but judge the clustering behavior according to the standard according to the result of clustering.Since the data obtained in reality are basically unlabeled data,clustering research is very necessary and important.From the perspective of the clustering method,the data samples need to be characterized before being clustered,which requires preprocessing the data.Deep neural networks(DNN)can transform data into a friendly representation,which has the inherent characteristics of a highly nonlinear transformation.Combining deep learning and clustering has become a new clustering method research-deep clustering.In this paper,we mainly study the convolutional neural network clustering method in the deep clustering method.We have studied convolutional neural network clustering methods from three aspects:network composition,clustering method and kernel method in deep clustering.We compared them with existing deep clustering methods.The main research work and results are as follows:(1)In order to deal with the problem that the initial clustering false label leads to the development of a worse method,we propose a convolutional neural network clustering algorithm based on multi-task learning.In the existing deep clustering method,because the label data set is needed for network training and learning,the label of the initial clustering result is used as the training sample.But the label may be wrong in itself,which leads to clustering in the wrong direction.In order to deal with it,we introduce a multi-task learning method.It makes the classification method and the clustering method guide each other to learn,and share the network parameters of the convolutional neural network.In addition,we also added a threshold to conditionally filter the clustering results to reduce the negative impact of wrong labels on clustering.The experimental results of the algorithm on the image data set prove that the method we propose is indeed beneficial to improve the clustering effect.(2)In neural networks,models are often only used for feature extraction,and the specific features are not explained.The features extracted by the network are still likely to be affected by the manifold space,and the features may still exist in the manifold structure.To solve this problem,we choose the clustering method of geodesic linear density peak,which can better reflect the manifold structure inherent in the data set.After that,we apply the clustering method to the model of deep clustering.The clustering method based on geodesic peak density is used to replace the traditional clustering method k-means.Finally,experiments are carried out on the image data set.The experimental results prove that the method we proposed does have the ability to deal with the problem of the manifold structure of image features.(3)This chapter proposes a clustering algorithm for convolutional neural networks based on kernel functions.In this chapter,we introduce kernel functions and kernel k-means clustering algorithm.Then we bring the idea of the kernel method into the method of deep clustering,hoping to establish the relationship between feature points,the relationship between images,by kernel method.Applying this method to image data,the experimental results prove that our proposed method is indeed helpful to improve the clustering effect.