DynMLP-GNN : A Dynamic Sampling Graph Neural Network with Multi-Layer Perceptron for Knowledge Graph Reasoning

Knowledge graphs have emerged as a pivotal area of research because of their extensive applications in personalized recommendations, drug interaction predictions, and intelligent question-answering systems. The existing methods based on graph neural networks use fixed propagation methods in knowledge graph inference, which often leads to information overload and reduces inference efficiency. To address this issue, this study introduces a novel graph neural network model, referred to as DynMLP-GNN, to enhance entity representations and improve path selection in knowledge graph reasoning. Specifically, this study incorporates an enhanced message-passing mechanism via multilayer perceptron, proposes a dynamic sampling strategy to prioritize entities most relevant to the query during propagation, and employs LSTM gating units to capture long-distance dependencies. Experimental results on multiple public knowledge graph datasets reveal that DynMLP-GNN substantially outperforms existing approaches in terms of both efficiency sand accuracy, particularly in addressing complex reasoning tasks.