Removal Experiments on Character Interaction Networks

Does network centrality identify protagonists? Franco Moretti (2011) argued that the protagonist is not the most connected character, but the one whose removal most destabilizes the network. We test this claim computationally on Game of Thrones — both the book series and the HBO show — using Andrew Beveridge's character interaction datasets.

Research Questions

1. Degree vs. destabilization: Do the most connected characters coincide with the most structurally critical ones? Or do centrality metrics and removal-based fragmentation produce different rankings?
2. Robustness to extraction method: The book dataset uses crude text co-occurrence (names within 15 words). The TV dataset uses richer interaction types (speech, mentions, shared scenes). If both produce similar destabilization rankings, the structural signal is robust to how the network is built.
3. Communities as storylines: Do Beveridge's detected communities correspond to known storylines (Stark arc, Lannister arc, Wall, Essos)?

Data

Character interaction networks from mathbeveridge/asoiaf (books) and mathbeveridge/gameofthrones (TV show). See data/README.md for provenance and licensing.

Quick Start

pip install -r requirements.txt
python scripts/download_data.py
python scripts/removal_experiment.py data/raw/asoiaf/asoiaf-book1-edges.csv results/book1_removal_ranking.csv

Structure

scripts/          Core algorithms (no plotting, no Jupyter dependency)
notebooks/        Reproducible analysis with narrative commentary
data/raw/         Unmodified source data
results/          Final tables
figures/          Publication-ready plots

References

• Moretti, F. (2011). "Network Theory, Plot Analysis." New Left Review, 68.
• Beveridge, A. & Shan, J. (2016). "Network of Thrones." Math Horizons, April 2016.

License

CC BY-NC-SA 4.0 (following Beveridge's dataset license).