Dual Mamba for Node-Specific Representation Learning: Tackling Over-Smoothing with Selective State Space Modeling

The introduction of the Dual Mamba-enhanced Graph Convolutional Network (DMbaGCN) marks a significant advancement in tackling the over-smoothing problem prevalent in deep Graph Neural Networks (GNNs). Over-smoothing occurs when repeated message passing leads to indistinguishable node representations, a challenge that existing solutions have only partially addressed. DMbaGCN innovatively combines two modules: the Local State-Evolution Mamba (LSEMba) focuses on local neighborhood aggregation, while the Global Context-Aware Mamba (GCAMba) incorporates global context through attention mechanisms. This dual approach not only enhances node discriminability but also allows for a more nuanced understanding of how node representations evolve across layers. The effectiveness of DMbaGCN has been validated through extensive experiments on multiple benchmarks, showcasing its potential to improve representation learning in GNNs significantly.