Method for Explaining Regression Prediction Results Using Machine Learning on Temporal Graph Data

Takumi HANANO, Taku HARADA

doi:10.34385/proc.95.31

説明

Understanding the temporal graph elements that contribute to a prediction is essential for improving interpretability in temporal graph neural networks (TGNNs). The existing explanation methods for graph neural networks have been proposed, such as GNNExplainer and PGExplainer, but these approaches mainly focus on static structures and often overlook time factors. As a result, they cannot identify the time intervals contribute that to model predictions. In this study, we propose a novel method that computes the importance of time intervals for each edge in a temporal graph. The proposed method is based on masking: edge values at consecutive time steps are hidden, and the resulting change in predictive performance is measured.

収録刊行物

IEICE Proceeding Series

IEICE Proceeding Series 95 250-258, 2025-12-03

The Institute of Electronics, Information and Communication Engineers

キーワード

詳細情報詳細情報について

CRID: 1390025650674484864

DOI: 10.34385/proc.95.31

ISSN: 21885079

本文言語コード: en

データソース種別

JaLC

抄録ライセンスフラグ: 使用不可

書き出し

問題の指摘