Pausanias Analysis - MANTO Place Survival Model

MANTO Place Survival Model

Can the mythological position of a place, using evidence available before Pausanias, predict whether Pausanias later describes it as surviving or as ruined, abandoned, deserted, or destroyed?

Leakage correction. Earlier experimental runs used Pausanias mention and passage counts as a “fame” baseline. Those measurements come from the same text that supplies the outcome label, so they have been removed. Earlier runs also allowed spelling aliases for the same MANTO entity to enter different folds; the corrected model uses exactly one row per MANTO place. This page excludes both sets of retired runs.

255linked labelled places

204survives

51does not survive

0.564best balanced accuracy

Model Comparison

The best current model improves balanced accuracy by +0.083 over the strictly pre-Pausanias attestation-volume baseline. This is a weak predictive result, not a practically reliable classifier. It is published here to make the present evidential limit visible.

The label-link join initially produced 444 alias rows. These were collapsed to 255 unique MANTO places, removing 189 duplicate spelling/link rows before cross-validation.

Results are pooled out-of-fold predictions from stratified five-fold cross-validation. Balanced accuracy averages recall for the two classes and is the primary measure because the majority-class ordinary accuracy is 0.800. An always-survives classifier has balanced accuracy 0.500.

Feature family	Balanced accuracy	Ordinary accuracy	Survivor recall	Non-survivor recall
Pre-Pausanias attestation	0.480	0.604	0.686	0.275
Connectedness best balanced accuracy	0.564	0.678	0.755	0.373
Connectedness + attestation	0.561	0.675	0.750	0.373
Network position	0.473	0.545	0.593	0.353
Network + connectedness	0.529	0.659	0.745	0.314
Geography	0.485	0.518	0.539	0.431
All admissible features	0.522	0.647	0.730	0.314

Best Model: Pooled Out-of-Fold Predictions

		Predicted
		Survives	Does not survive
Actual	Survives	154	50
Actual	Does not survive	32	19

What the Model Is Allowed to See

Pre-Pausanias Attestation

Raw MANTO edge volume from evidence strictly earlier than Pausanias. It contains no counts derived from Pausanias.

Network Position

Degree, PageRank, k-core, community role, reach, redundancy, bridges, and paths to major places in the earlier myth network.

Connectedness

Shared figures and action patterns, distinctive local figures, kin-mediated ties, large-place connections, and temporal layering.

Geography

Distances, nearby-place counts, and whether narrative neighbours are geographically local.

Strongest Standardized Coefficients

The final logistic regression is fitted on standardized features. Coefficients show conditional associations within a correlated feature set; they are not causal effects and should be interpreted as patterns rather than isolated rankings.

Feature	Family	Coefficient	Direction
Kin-Linked Neighbor Count	Connectedness	-1.275	does not survive
Shared-Mythic-Figure Neighbor Count	Connectedness	+1.040	survives
Shared-Figure Large-Place Neighbor Count	Connectedness	+0.852	survives
Shared Figure Count Z-Score	Connectedness	-0.813	does not survive
Shared-Figure Neighbor Z-Score	Connectedness	+0.721	survives
Action Pattern Count	Connectedness	+0.714	survives
Early Imperial Attestation Count	Connectedness	-0.696	does not survive
Exclusive Figure Count	Connectedness	+0.695	survives
Kin-Linked Large-Place Count	Connectedness	+0.619	survives
Local Place Neighbor Count	Connectedness	-0.583	does not survive
Shared-Action Neighbor Count	Connectedness	-0.555	does not survive
Mean Figure Ubiquity	Connectedness	-0.532	does not survive
Shared-Action Large-Place Neighbor Count	Connectedness	-0.531	does not survive
Maximum Figure Ubiquity	Connectedness	+0.493	survives
Archaic Attestation Count	Connectedness	+0.474	survives
Earliest Attestation Year	Connectedness	-0.464	does not survive

Browse definitions and examples for all 41 connectedness features.

Scope and Remaining Checks

The target is what Pausanias reports in the second century, not modern archaeological survival.
The active labels are LLM-extracted and still require a stratified human gold-set audit.
The model uses exact-name and deterministic transliteration links; unreviewed LLM-curated links are excluded.
All spelling aliases are collapsed to one MANTO entity before folds are assigned.
Cross-validation is stratified by label, but leave-one-region-out validation has not yet been run.
Performance supports association and prediction within this dataset, not a causal claim about why settlements survived.

The place-state sweep reviewed 3,170 of 3,170 passages; 1,098 passages contain at least one explicit place-state claim. MANTO release 19446255 (10.5281/zenodo.19446255); model-facing MANTO edges are restricted to evidence dated strictly before Pausanias.

Inspect the MANTO place network · Inspect Pausanias–MANTO place links