Research on the Influence of Modified Activation Function on the Graph Classification Model

In recent years, graph data has become increasingly prevalent in the actual world and can be utilized to represent the intricate link between each entity. To solve the problem of graph classification in graph neural networks, the effect of verification on the classification accuracy of various models based on modified activation functions is proposed using the IMDB-BINARY and PROTEINS datasets as the experimental dataset with graph convolutional network, graph sampling network, and other models to conduct experiments; keeping the dataset unchanged, the classification accuracy of each model was determined and compared by modifying the activation function of the model and the relevant parameters. Following testing, the activation function has a significant impact on the categorization of graph data in the graph classification problem. The ReLU function is the most applicable in the classification of the selected models in this study, and the graph convolution model is the optimal model for addressing such challenges.