Graph Learning for Complex Medical Data Integration and Predictive Inference

Knowledge graph learning frameworks currently confront two significant challenges: Firstly, in multi-relational heterogeneous graphs, it remains difficult to simultaneously evaluate the effects of nodes and relationships on graph representations. Secondly, the issue of overcoming the constraints associated with first-order subgraphs during aggregation, as well as the construction of more task-specific multi-hop domain subgraphs, has yet to be adequately addressed. This research mitigates these challenges by utilizing subgraph aggregation techniques within the context of knowledge graph learning. The proposed model is refined to enhance the representation learning of multi-relational heterogeneous graphs. At the one-hop neighbor level, the model integrates the Relational Graph Convolutional Network (RGCN) with an attention mechanism, akin to the enhanced Graph Attention Network (GAT), allowing it to concurrently consider the impact of neighboring nodes and edges on the central node’s representation, as well as the central node’s significance when one-hop neighbor subgraphs are consolidated. At the multi-hop neighbor level, the model extends its reach to encompass nodes at multiple hop distances, employing a gating mechanism to extract relevant information. This approach circumvents the limitations of one-hop neighborhood subgraphs by adopting multi-hop neighbors as the neighborhood subgraph, thereby enabling the gating mechanism to adaptively represent the central nodes. The effectiveness of the enhanced model is empirically validated on a single-relation homogeneous graph, demonstrating notable performance improvements on the Cora dataset and the multi-relation heterogeneous FBISK-237 dataset across several metrics, including Accuracy, Mean Rank (MR), Mean Reciprocal Rank (MRR), and Hit@n. Additionally, ablation studies corroborate the effectiveness of each module within the model through various validation methods.