A naukri.com initiative
Medium
2M
326
Image Credit: Medium
Tracing Historical Political Leaning in Newspapers with GNNs
- The political leanings of newspapers across the United States have long been a focus of social science research.
- Researchers sought answers to questions about change of newswire activity across article topics and decades, and how it relates to polarization of that topic in the United States.
- Dataset from Harvard Professor Melissa Dell’s lab on newspaper articles from 1878 to 1977 was explored to answer these questions.
- Graph Neural Network (GNN) architectures were used to generate rich embeddings that can capture newspaper similarity.
- Models were trained and compared to predict political ideology based on input features (metadata and node embeddings).
- Results showed that the heterogeneous SAGEConv model had the highest test performance compared to the other models.
- The models' representative node embeddings were then analyzed to group similar outlets together and identify relationships between them.
- Future analysis should incorporate additional features, such as income and demographic data for more accurate GNN model building.
- Overall, the study revealed that graph embeddings may not provide a distinct advantage over metadata itself.
- Future work should explore dynamic graph neural networks that explicitly model temporal dependencies.
Read Full Article
19 Likes
For uninterrupted reading, download the app