Graph Neural Networks for Large-Scale Imbalanced Data

2022 Fiscal Year Final Research Report

• Project/Area Number: 21K21280
• Research Category: Grant-in-Aid for Research Activity Start-up
• Allocation Type: Multi-year Fund
• Review Section: 1001:Information science, computer engineering, and related fields
• Research Institution: The University of Tokyo
• Principal Investigator: Suzumura Toyotaro 東京大学, 情報基盤センター, 教授 (70552438)
• Project Period (FY): 2021-08-30 – 2023-03-31
• Keywords: グラフニューラルネットワーク / 機械学習 / 人工知能 / グラフ解析 / 大規模データ / 高性能計算

Outline of Final Research Achievements

In this study, we proposed the optimization of a GNN (Graph Neural Network) model and software architecture to capture the characteristics of graph data with imbalanced labels and long-term dynamic graph data, which are used as data representations in real-world applications. For imbalanced labels, we proposed a method to construct a heterogeneous graph and train it with a heterogeneous GNN model, which achieved high performance in predicting fraudulent accounts in financial transaction networks. Additionally, we proposed the Spectral Wavelet, which captures the characteristics of dynamic large-scale graphs in a long-term context, and an efficient method for evaluating the complementarity of knowledge graph relationships using topological data analysis.

Free Research Field

人工知能

Academic Significance and Societal Importance of the Research Achievements

大規模グラフデータを扱う機械学習などの研究では、動的なグラフ構造の変化、インバランスなラベルが要因となるモデル性能の問題を解決することは必須の課題である。本研究と並行して企業（自動車会社、新聞社、人材紹介会社）との共同研究を進めていく中でもこれらの課題が本質的な課題であることを確認しており、学術的な意義ばかりではなく社会的にも大きな意義のある研究成果と言える。