Disentangling Political Bias in Social Media Data Through Multitask Embedding Learning

Yihong Zhang, Takahiro Hara, Lina YaoIEEE Transactions on Computational Social Systems 2025 · DOI

Abstract

Socialmedia is an important information source for understanding community events, personal communications, and other social activities. Therefore, mining information from social media data has been widely studied. Recently, it has been shown that knowledge graphs constructed from social media data can be effectively mined, which usually results in embeddings that contain structural information about social media entities. However, it has been shown that embeddings learned from social media data can contain social bias. Bias in the embedding is usually considered as undesirable, but there is only limited number of works that can separate bias from the embedding. In this article, we propose a bias disentanglement model that separates embedding into two parts, one preserves general information but has bias removed, and the other contains distilled bias. The two sets of embeddings can be used for computational tasks with different information requirements. Our method, disentangling bias through multitask embedding learning (DisMEL), is based on knowledge graph embedding techniques and supervised disentangling models. In addition to qualitative analysis of the generated embeddings, we test our method in two downstream tasks, namely, user political bias prediction and tweet recommendation. We compare our method with several baseline methods in experimental evaluations, based on a large-scale Japanese political social media dataset we collected and prepared. We show that when the debiased embedding set is used, the tasks irrelevant to political bias can be improved. On the other hand, when the biased embedding set is used, the tasks relevant to political bias can be improved. While we deal with political bias in this article, our method can also be adopted for other kinds of bias, such as gender or ethnicity bias.
BibTeX
@article{zhang2025disentangling,
  author = {Yihong Zhang and Takahiro Hara and Lina Yao},
  title = {{Disentangling Political Bias in Social Media Data Through Multitask Embedding Learning}},
  journal = {IEEE Transactions on Computational Social Systems},
  year = {2025},
  doi = {10.1109/tcss.2025.3576342},
}

Loading related content...

Figure