Why some judgments matter more than others

Research question

Citation counts flatten a judgment’s position in the wider legal network. This study tests when composite centrality measures outperform their individual components across unbalanced and ground-truth-balanced ECHR citation subnetworks.

The methodological boundary

The repository has two supported depths. make from-metadata rebuilds citation edges and article-edge networks from the canonical ECHR metadata export. The paper-facing analyses deliberately start later, at fixed centrality-enriched total_df.csv bundles, so make paper-results can regenerate the statistical tables, threshold tests, network-family comparisons, priority comparison, and figures without recomputing every centrality result.

That checkpoint is an explicit reproducibility decision: expensive network construction remains inspectable, while the claims made in the manuscript can be rerun quickly from stable inputs.

What the comparisons found

Across 66 subnetworks, Degree + Eigenvector composites beat their individual components in 45 networks against the importance ground truth and 43 against doctypebranch. Degree + InDegree won 45 of 66 under both ground truths. The optimized Degree + Eigenvector weighting changed from 0.30 to 0.35 between those ground truths, evidence that the ranking is contingent on what “important” is taken to mean.

The alternative that prioritised lower-relevance cases performed better in only 1 of 18 comparisons; the higher-relevance approach performed better in 17, with an average difference of −31.6% for the low-relevance variant.

Traceability and contribution

I worked across the research code and its reproducibility layer: metadata processing, citation-edge and network generation, centrality experiments, comparison scripts, result packaging, and the claim-to-artifact map. Each retained quantitative claim is paired with the script that produces it and the concrete CSV, JSON, or figure output that supports it. Active scripts live under scripts/; notebook-era and superseded experiment families are isolated under archive/ rather than left in the executable path.

Publication state

The paper, by Gijs van Dijck and David Henry Francis Wicker, is accepted for ICAIL 2026. Final proceedings publication is pending and no DOI has been assigned yet. The complete citation, DOI, and public manuscript link will be added when released.

Interpretation limits

Centrality does not establish an objective legal importance ranking. The result depends on network construction, the selected ground truth, the measures combined, and the threshold used to combine them. The contribution is therefore not a universal winner; it is an inspectable account of how those decisions change which judgments move from peripheral to central.

@inproceedings{vandijckwicker2026centrality,
  title = {Why Some Judgments Matter More Than Others: Advancing Network Centrality Analysis},
  author = {van Dijck, Gijs and Wicker, David Henry Francis},
  booktitle = {Proceedings of the International Conference on Artificial Intelligence and Law},
  year = {2026},
  note = {Accepted; final proceedings metadata and DOI pending}
}

TY  - CPAPER
TI  - Why Some Judgments Matter More Than Others: Advancing Network Centrality Analysis
AU  - van Dijck, Gijs
AU  - Wicker, David Henry Francis
T2  - International Conference on Artificial Intelligence and Law
PY  - 2026
N1  - Accepted; final proceedings metadata and DOI pending
ER  -