Abstract
With the emergence of graph neural networks (GNNs), deep learning techniques for non-Euclidean data have become the go-to method for various graph processing tasks. However, many GNNs suffer from over-smoothing, in which node features become indistinguishable with the use of multiple message passing layers and do not generalize real-world non-homogenous graph data well. To address these issues, we propose an enhanced adjustable method that attends to important hops via independent learnable weights and includes an initial connection method to further stabilize the larger model. We conducted extensive experiments on 12 real-world graph benchmark datasets of various sizes and network homophily levels to show our approach outperforms several recently proposed adaptive methods for node classification tasks.
Original language | English |
---|---|
Pages (from-to) | 435-452 |
Number of pages | 18 |
Journal | Information Sciences |
Volume | 613 |
DOIs | |
Publication status | Published - 2022 Oct |
Keywords
- Attention
- Deep learning
- Graph neural networks
- Homophily
- Over-smoothing
ASJC Scopus subject areas
- Software
- Information Systems and Management
- Artificial Intelligence
- Theoretical Computer Science
- Control and Systems Engineering
- Computer Science Applications