2021 Fiscal Year Research-status Report
Neural Network based Graph Learning: Model Evolution and Real-World Application
| Project/Area Number |
21K12042
|
|---|
| Research Institution | National Institute of Advanced Industrial Science and Technology |
|---|
Principal Investigator |
劉 欣 国立研究開発法人産業技術総合研究所, 情報・人間工学領域, 主任研究員 (20803935)
|
|---|
| Project Period (FY) |
2021-04-01 – 2024-03-31
|
| Keywords | Graph Neural Network / Graph Embedding / Geometric Learning / Graph Mining / Node Classification / Link Prediction |
|---|
| Outline of Annual Research Achievements |
In terms of model evolution, we have proposed GNNs for heterogeneous graphs (multiple types of nodes and edges). Our approaches capture the rich semantic information and thus significantly outperform SOTA methods in link prediction, node classification, and node clustering tasks.
In terms of real-world application, we have proposed GNN-based approaches for various applications: a) We propose a GNN-based neural recommender method to solve the temporal dynamics problem. b) We propose Spatio-Temporal-Categorical GNN to solve the fine-grained incident prediction problem. c) We develop a citation network embedding algorithm to enhance the citation recommendation performance. d) We design a GNN-based model to handle cross-lingual text classification. In all of these applications, we demonstrate that our new approaches are superior to existing ones.
|
| Current Status of Research Progress |
Current Status of Research Progress
1: Research has progressed more than it was originally planned.
Reason
The current progress of the project is going well. There are no delays. The finished work has been published in top conferences and journals. We are working now on heterogeneous graph embedding with single-level aggregation instead of the traditional bi-level aggregation. We have finished some of the work and submitted the papers.
|
| Strategy for Future Research Activity |
We will focus on the following problems: 1) Feature selection in GNNs, 2) GNN for imbalanced label classification, 3) GNN for temporal graphs.
We will also continue develop GNN-based approaches for solving various practical problems and applications.
|
| Causes of Carryover |
We planned to buy a workstation and a graphics card in FY2021. However, the product release date of the fittest models was delayed to May/June 2022. We will buy them in FY2022.
|
