Research On Classification Of Imbalance Data Based On Graph Neural Network

Traditional machine learning methods and deep learning methods have been widely used in classification tasks,such as random forests,support vector machines,and neural networks.which assume independent and homogeneous distribution among samples when training,but in the real world,everything in the world is potentially connected,and graph data can represent this information adequately,so graph neural networks(GNNs)for graph data classification have emerged.Graph Neural Network(GNN)model for graph data classification was born.In practical applications,class imbalance is common,for example,in online comment data,the amount of fake comments and real comments is unbalanced.Although existing graph neural network classification models have achieved great success in classification tasks,they can only show good classification results when the number of samples in each class is not very different,so it is very important to study graph neural networks for unbalanced node classification.important.There are many real-world data that are unbalanced,and to address this problem,this paper proposes an integrated graph neural network model for unbalanced node classification.Using a spectral domain-based graph convolutional neural network as the base classifier,multiple base classifiers are trained in parallel,and then the predictions of multiple base classifiers are aggregated using the Bagging integrated learning method,and finally the final classification result is determined using the majority voting method.The differences of base classifiers mainly come from the differences of training sets.Each training set is randomly sampled from the original data set.Samples of various types are sampled according to the experimental setting,and then most samples are randomly downsampled according to the unbalance ratio,which is combined with a few samples to form a balanced training set.To address the problem that false comment detection is difficult and its data is unbalanced,this paper proposes a false comment detection model based on Graph Convolutional Network(GCN).Each comment is treated as a node,and the sub-graphs constructed based on three different relationships are aggregated to form a total graph,after which the constructed graph data are fed into the classification model proposed in this paper for training.Firstly,real comments and false comments are sampled According to the experimental setup,followed by random undersampling of real comment samples According to the imbalance ratio(normal comment samples that are not sampled are not discarded,and their node features are also involved in model training,only as unlabeled data for semi-supervised training),and the sampled real comment samples and extracted fake comment samples were combined into a balance training set.The steps were repeated,and different training sets were input into each graph convolutional neural network for training.Finally,the predicted results of all base classifiers were integrated to determine the final output results.The experimental results show that the proposed model is effective in unbalanced node classification and false comment detection.