A new link prediction model for grain trade networks based on improved variational graph autoencoder and genetic algorithm

Food security is related to the national economy and people’s livelihood, and exploring potential cooperative relationships with oneself is one of the most effective strategies to prevent and mitigate the risk of food import supply chain disruptions. How to use prior information in the grain import trade network to obtain better potential representations of nodes is a key issue in link prediction tasks. Under the theoretical framework of the variational graph autoencoder, this paper creates a new link prediction model, IVGAE-GA. Two feature extraction modules and a feature fusion module are designed to mine effective information in the trade networks. Specifically, a dynamic adaptive graph attention (DAGAN) module is proposed to extract high-order feature information from trade networks. Then, the neighborhood feature information of each node is captured through the graph convolutional neural network (GCN) to strengthen the guiding effect of the initial prior information on the prediction results. In addition, an average feature fusion (AVFF) module is designed to further refine the latent representation of nodes by mixing these non-local and local feature information. The entire IVGAE framework is optimized through cross-entropy loss and KL loss. Finally, the genetic algorithm (GA) is utilized for hyperparameter selection to help the model perform better. Extensive experimental results on two widely used publicly available datasets and four real grain trade networks illustrate that our model achieves better prediction performance compared to some existing methods. The proposed link prediction framework can be a good option for predicting potential cooperative relationships.