Meta-Learning Based Graph Neural Networks for Online Discussions
Meta-Learning Based Graph Neural Networks for Online Discussions


In order to achieve large-scale consensus support among topic-based online discussions, we have developed a crowd-scale discussion support system called D-Agree. It is based on agents that can automatically facilitate users’ arguments such as classifying labels (e.g., issues and ideas ) of those arguments during discussions. Despite the recent success of applying graph neural networks (GNN), a graphical deep learning based method, for facilitating arguments on D-Agree, generalization of cross-topic remains an open problem. This is because the trained GNN models based on existing topics might completely fail in the facilitation of new topics. To address this issue, we propose a three-stage pattern-based meta-GNN method that 1) first learns a high-level pattern representation in argument data. 2) It then utilizes a meta-learning framework to train a series of specific topic-based GNN models based on the pattern knowledge. 3) In the third stage, a meta-learner is trained with considering all specific topic-based GNN models and allows it to generalize to new topics quickly.


Consensus support, Online discussion system, Text classification, etc.


氏名 専攻 研究室 役職/学年
丁 世堯 社会情報学専攻 伊藤研 助教
伊藤 孝行 社会情報学専攻 伊藤研 教授