講演情報
[1K-05]HashtagMeta: Fake News Mitigation for Hashtag Recommendation
*史 家梁1、駒水 孝裕1、井手 一郎1 (1. 名古屋大学 大学院情報学研究科)
発表者区分：学生
論文種別：ロングペーパー
インタラクティブ発表：あり
論文種別：ロングペーパー
インタラクティブ発表：あり
キーワード：
推薦システム、ソーシャルメディア
Hashtag recommendation systems facilitate user engagement in discussions on social media, but they also contribute to the propagation of fake news by inadvertently promoting misleading hashtags. To address this issue, we propose HashtagMeta, a novel graph neural network based framework designed to reduce the spread of fake news hashtags
. HashtagMeta is based on the information network consisting of tweets and hashtags, where tweet-hashtag relationship represents the containment of hashtags in each tweet.
It estimates the relevant hashtags to a tweet based on the connectivity in the network, which is realized via a graph neural network model. Our approach dealing with mitigating fake news consists of two main steps: (1) an unsupervised phase where fake news tweets are removed, node similarity method PathSim calculated which model semantic relationships in true tweets, and hashtag-hashtag edges based on the similarities are added to enhance connectivity of the network to deal with the data sparcity issue; and (2) a supervised phase leveraging Heterogeneous Graph Attention Networks to learn hashtag embeddings and minimize the propagation likelihood of fake news. Experimental evaluation demonstrates that our method effectively mitigates misleading hashtags while retaining core semantic information, providing a robust solution to misinformation and establishing a foundation for future graph-based research on fake news mitigation.
. HashtagMeta is based on the information network consisting of tweets and hashtags, where tweet-hashtag relationship represents the containment of hashtags in each tweet.
It estimates the relevant hashtags to a tweet based on the connectivity in the network, which is realized via a graph neural network model. Our approach dealing with mitigating fake news consists of two main steps: (1) an unsupervised phase where fake news tweets are removed, node similarity method PathSim calculated which model semantic relationships in true tweets, and hashtag-hashtag edges based on the similarities are added to enhance connectivity of the network to deal with the data sparcity issue; and (2) a supervised phase leveraging Heterogeneous Graph Attention Networks to learn hashtag embeddings and minimize the propagation likelihood of fake news. Experimental evaluation demonstrates that our method effectively mitigates misleading hashtags while retaining core semantic information, providing a robust solution to misinformation and establishing a foundation for future graph-based research on fake news mitigation.