File Download
There are no files associated with this item.
Links for fulltext
(May Require Subscription)
- Publisher Website: 10.1109/IJCNN54540.2023.10192005
- Scopus: eid_2-s2.0-85169591777
- WOS: WOS:001046198707072
Supplementary
- Citations:
- Appears in Collections:
Conference Paper: HES: Edge Sampling for Heterogeneous Graphs
Title | HES: Edge Sampling for Heterogeneous Graphs |
---|---|
Authors | |
Keywords | Edge Sampling Graph Neural Networks |
Issue Date | 18-Jun-2023 |
Abstract | In light of the success of graph neural networks (GNNs), recent years have seen significant developments in modeling graphstructured data. Heterogeneous graphs have been widely adopted to model complex systems for various ML tasks. However, although some researchers have proposed methods for heterogeneous graphs, they merely focus on node features while neglecting the effectiveness of edge features. Besides, some research projects attempted incorporating edges into GNNs, but most regarded edge features as shared weights between node pairs. In this paper, we propose a two-stage method HES to learn the correlations among edge neighbors: (1) Graph Transformation: We convert the original heterogeneous graph into an undirected graph while preserving the orientation information, (2) Group Edge Sampling: To reduce the computation cost for the edge sampling in a heterogeneous graph, we propose to sample the most important edges over a group of edge neighbors instead of the whole graph, which leverages the edge features based on the one-hot encodings to describe the mutual influences between any adjacent edges. Finally, the experimental results on multiple public datasets show that HES outperforms existing state-of-the-art (SOTA) graph sampling methods. We further apply our approach to some existing GNN models as a pre-training process, demonstrating that HES can augment GNN-based models effectively. |
Persistent Identifier | http://hdl.handle.net/10722/333887 |
ISI Accession Number ID |
DC Field | Value | Language |
---|---|---|
dc.contributor.author | Fang, Le | - |
dc.contributor.author | Wu, Chuan | - |
dc.date.accessioned | 2023-10-06T08:39:54Z | - |
dc.date.available | 2023-10-06T08:39:54Z | - |
dc.date.issued | 2023-06-18 | - |
dc.identifier.uri | http://hdl.handle.net/10722/333887 | - |
dc.description.abstract | <p>In light of the success of graph neural networks (GNNs), recent years have seen significant developments in modeling graphstructured data. Heterogeneous graphs have been widely adopted to model complex systems for various ML tasks. However, although some researchers have proposed methods for heterogeneous graphs, they merely focus on node features while neglecting the effectiveness of edge features. Besides, some research projects attempted incorporating edges into GNNs, but most regarded edge features as shared weights between node pairs. In this paper, we propose a two-stage method HES to learn the correlations among edge neighbors: (1) Graph Transformation: We convert the original heterogeneous graph into an undirected graph while preserving the orientation information, (2) Group Edge Sampling: To reduce the computation cost for the edge sampling in a heterogeneous graph, we propose to sample the most important edges over a group of edge neighbors instead of the whole graph, which leverages the edge features based on the one-hot encodings to describe the mutual influences between any adjacent edges. Finally, the experimental results on multiple public datasets show that HES outperforms existing state-of-the-art (SOTA) graph sampling methods. We further apply our approach to some existing GNN models as a pre-training process, demonstrating that HES can augment GNN-based models effectively.</p> | - |
dc.language | eng | - |
dc.relation.ispartof | the International Joint Conference on Neural Networks (18/06/2023-23/06/2023, Queensland) | - |
dc.subject | Edge Sampling | - |
dc.subject | Graph Neural Networks | - |
dc.title | HES: Edge Sampling for Heterogeneous Graphs | - |
dc.type | Conference_Paper | - |
dc.identifier.doi | 10.1109/IJCNN54540.2023.10192005 | - |
dc.identifier.scopus | eid_2-s2.0-85169591777 | - |
dc.identifier.volume | 2023-June | - |
dc.identifier.isi | WOS:001046198707072 | - |