Graph Neural Networks (GNNs) have shown remarkable success in graph representation learning, but often struggle to capture both local and global structural information effectively. While recent approaches like LGL-GNN incorporate fixed motifs to enhance local feature extraction, they suffer from domain dependency and suboptimal motif selection. In this paper, we propose AdaMotif-GNN, a novel framework that dynamically selects and weights motifs based on graph characteristics and task requirements. Our method introduces a motif attention mechanism that learns to prioritize different motif types during training, eliminating the need for manual motif selection. Extensive experiments on benchmark datasets demonstrate that AdaMotif-GNN outperforms state-of-the-art methods, achieving average improvements of 2.3% on bioinformatics datasets and 1.8% on social network datasets compared to fixed-motif approaches.
Graph Neural Networks (GNNs) have shown remarkable success in graph representation learning, but often struggle to capture both local and global structural information effectively. While recent approaches like LGL-GNN incorporate fixed motifs to enhance local feature extraction, they suffer from domain dependency and suboptimal motif selection. In this paper, we propose AdaMotif-GNN, a novel framework that dynamically selects and weights motifs based on graph characteristics and task requirements. Our method introduces a motif attention mechanism that learns to prioritize different motif types during training, eliminating the need for manual motif selection. Extensive experiments on benchmark datasets demonstrate that AdaMotif-GNN outperforms state-of-the-art methods, achieving average improvements of 2.3% on bioinformatics datasets and 1.8% on social network datasets compared to fixed-motif approaches.
