A dynamic recommendation method based on an improved graph neural network is proposed to address the problem of excessive target resource demand nodes， which result in low adaptability between learners and virtual simula-tion experiment resources， and poor recommendation performance. A multidimensional feature association graph of experiment resources is constructed and modify the graph neural network architecture is modified to facilitate the propagation of information regarding target resource demand nodes within the association graph. The Sigmoid activation function is used to extract feature information of virtual simulation experiment resources from this association graph. Based on this， a dynamic recommendation architecture based on the improved graph neural network is constructed to generate initial low-dimensional embedding vectors， construct first-order collaborative signals propagated through interaction behavior， update the embedding vectors in real time， and calculate the compatibility between any learner and an experiment resource， using this compatibility as the basis for recommendations. To validate the effectiveness of the method， four typical learners （numbered 1，2，3， and 4） were selected for testing. The experimental results show that this method recommends resources with the highest adaptation rates of 40%，80%，60%， and 80% for learners 1，2，3， and 4， respectively， which is consistent with the experimental indicators， and takes only 4 seconds to obtain the best recommendation effect.