| In recent years, Graph Neural Networks (GNNs) have made significant progress in learning representations on static, isomorphic graphs. However, real world graph data are often dynamic and heterogeneous, presenting new challenges. This study proposes a Dynamic Heterogeneous Graph model (DHGA) designed to address the evolving relationships and structural changes inherent in such graphs. The DHGA captures both spatial and temporal dependencies by modeling temporal evolution and aggregating relationships, effectively adapting to changes in graph entities over time and integrating information from neighbor entities. It preserves the diversity of graph data to provide accurate node representations. For each snapshot of the graph at a given time, the model aggregates information from neighboring nodes based on internal and external relationships, handling nodes of the same and different types accordingly. A transformer-based architecture is employed to process information across various temporal slices, enhancing the aggregation of temporal data. Additionally, the DHGA incorporates a temporal evolution module that tracks and adapts to changes in node features, facilitating better generalization over time and improving the modeling of trends in node evolution across consecutive graphs. A Mixture of Experts approach is utilized during the aggregation process to combine outputs from multiple models, thereby increasing prediction accuracy. The proposed model was evaluated on three dynamic, heterogeneous datasets characterized by evolving node features. Experimental results demonstrate that the DHGA outperforms traditional models in graph representation learning, showcasing its efficacy in dynamic environments. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.