Multimodal data multidimensional time series fusion method based on graph neural network

Intelligent algorithms are often limited in fusion accuracy in multi-modal data multi-dimensional time series fusion due to the semantic gap between modalities and insufficient coupling of dynamic time series. To break through this bottleneck, a multi-modal data multi-dimensional time series fusion method based on graph neural network (GNN) is proposed. The dynamic graph structure is constructed by initializing the Gaussian kernel metric-driven graph, a modality-specific encoder is designed to extract differentiated features, and the fusion of multi-modal data multi-dimensional time series is realized by using iteratively updated graph neural network. Experiments show that the research method makes the true positive rate of method 1 in the receiver operating characteristic curve (ROC curve) reach about 0.9 when the false positive rate is 0.2, which is about 0.2, 0.25, and 0.3 higher than methods 2, 3, and 4, respectively. The information completeness and feature diversity of modal complementary gain are increased by an average of +33.2%, +40.4%, and the comprehensive gain is +36.8%, all of which are ahead of other methods. Experiments show that the research method integrates the precise modeling of multimodal data structure and deep feature mining through graph neural network, which significantly improves the accuracy of multi-dimensional time series fusion of multimodal data.