A heterogeneous network-based on interactions between plants was constructed，and a graph neural network-based method for predicting interactions between plants was proposed to solve the problem of accurately predicting inhibitory and promotive interactions between plants based on prior knowledge.The method is primarily composed of three modules including a representation learner，an interaction identifier，and a type classifier.The representation learner is responsible for extracting representations of interactions.The interaction identifier collaborates with the representation learner to predict whether an interaction exists on types of targeted plant and applies perturbations to the type classifier to maximally acquire transferable features across different types of interaction.The type classifier is designed to accurately distinguish types of interactions between plants to avoid being affected by the interaction identifier.The inter adversarial relationship among the three modules is leveraged to enhance robustness against the differences in the types of inhibitory and promotive interactions between plants，thereby effectively solving the problem of accurately predicting targeted interactions between plants.The AUC，precision，and accuracy of this method on the dataset of companion plant improved by 7.74，1.61，and 8.62 per cent，respectively compared with the best method SVM that of existing methods including SEAL，GATNE，HeGAN，PME，SVM，and RF，reaching 92.00%，80.12%，and 86.21%，respectively.The results indicate that the proposed method effectively mitigates the interference caused by target interaction type differences，enabling accurate prediction of plant interactions.This approach can be applied to optimize agricultural production practices.