Research On Classification Algorithm Based On Representation Learning

The explosive growth of data generated by various ways such as the Internet makes data analysis extremely difficult.As a basic problem of data mining,classification has become a hot topic.With the continuous development of machine learning and deep learning,the performance of classification algorithms based on representation learning has been greatly improved.However,the current classification algorithm based on representation learning has the following problems:the spatial relationship between nodes is not considered when the training set is expanded,the performance of classification is poor when the data labels are unbalanced,and the relationship between nodes cannot be considered when multi-label classification is implemented.Therefore,the focus of this paper is to put forward effective solutions to the above mentioned problems,so as to get better classification performance.The main research contents of this paper are as follows:(1)This paper studies and analyzes the spatial relationship between nodes without considering the expansion of the training set,and proposes the ET-GCN algorithm.Firstly,the nodes in the network are pre-classified to get the pseudo-labels of nodes,and then the centers of different categories are obtained by clustering the pseudo-labels of nodes by using the distribution density relationship of nodes.K nodes that are closest to the category center are selected to be added into the training set for training.Experimental results show that the proposed method is more effective than the existing methods.(2)In this paper,the semi-supervised classification algorithm of graph convolutional neural network is analyzed and studied.In order to solve the problem that the label data imbalance is not considered when the graph convolutional neural network is doing the classification task,an enhanced classification GCNBoost algorithm is proposed in this paper.The algorithm is based on the classification error to increase the attention to the classification error node and improve the loss function of graph convolutional neural network.The experimental results show that the proposed method is more effective than the existing methods.(3)The multi-label classification algorithm based on feature aggregation is studied and analyzed,and the MLCAA algorithm is proposed.Firstly,the attribute graph is constructed according to the attribute relationship,and then the feature is aggregated according to the constructed attribute graph.In addition,in order to avoid the interference of noise,the features of nodes are extracted by the denoising autoencoder,and the aggregated features are input into the label prediction of data.The experimental results show that the proposed method is more effective than the existing methods.