International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

Dual-Mode Density-Aware Attention-based Hierarchical Graph Pooling

Authors :

Roya Booryaee¹ Parsa Haddadian² Ali Kamandi³

1- Department of engineering science, college of engineering, University of Tehran 2- Department of engineering science, college of engineering, University of Tehran 3- Department of engineering science, college of engineering, University of Tehran

Keywords :

Graph Representation Learning،Adaptive Graph Pooling،Attention Mechanism،Sparse Graphs،Personalized PageRank

Abstract :

Graph pooling is a crucial operation in Graph Neural Networks (GNNs) for enabling scalable and expressive graph-level representation learning. However, existing pooling methods often overlook graph density and apply uniform attention mechanisms regardless of structural variability. In this paper, we propose DA-MAGPool, a dual-mode density-aware attention-based hierarchical pooling framework that dynamically selects between sparse local attention and multi-hop Personalized PageRank-based attention, conditioned on the input graph’s density. By introducing a learnable threshold to distinguish between sparse and dense graphs, our method adaptively applies the most suitable attention strategy, thereby preserving important local and global structural information. A unified node scoring mechanism is then used for Top-k pooling to retain salient nodes. Experimental results on five benchmark datasets demonstrate that DA-MAGPool consistently outperforms existing pooling baselines on four datasets and ranks second on the remaining one, achieving high accuracy across a range of graph classification tasks.