Reports / external_padic_aggregation_experiments.md
External P-Adic Aggregation Experiments
Generated by scripts/run_external_padic_aggregation_experiments.py.
Datasets: prepared external document-tree corpora. Max rows: 60. Max features: 30. P-adic subset sizes: 1, 2.
Feature panels: coverage panel plus 0 sampled panels. P-adic random subsets per size: 6.
Headline
- Retained affine p-adic member rows whose nearest branch set includes
ROOT: 0.0000
- Retained affine p-adic member rows with nonzero y-intercepts:
0.9616
- Retained affine p-adic member rows whose raw prediction equals the y-intercept:
0.8569
- Datasets where some p-adic variant improves on the current medoid p-adic loss:
1/4
- Datasets where the best p-adic variant beats or ties the best logistic/tree/forest row:
3/4
- Summary rows with at least one tie-scored fold:
26
- P-adic fit requests skipped by the point-system cap:
4
P-Adic Versus Euclidean
| dataset |
current_padic_loss |
best_padic_policy |
best_padic_loss |
best_euclidean_group |
best_euclidean_loss |
best_padic_beats_or_ties_euclidean |
| catena_ephesians |
0.0000 |
no_intercept_deepest_member |
0.0000 |
Random forest |
0.0000 |
1 |
| catena_galatians |
0.0000 |
no_intercept_deepest_member |
0.0000 |
Random forest |
0.0000 |
1 |
| catena_romans |
0.0167 |
no_intercept_deepest_member |
0.0167 |
Random forest |
0.0167 |
1 |
| comp_tc_mark_stemma_backbone |
0.3167 |
drop_zero_members |
0.2000 |
Logistic |
0.1505 |
0 |
Best P-Adic Policies
| dataset |
padic_policy |
subset_size |
mean_padic_loss |
mean_hard_padic_loss |
mean_adjusted_parameters_used |
tie_folds |
fallback_folds |
| catena_ephesians |
no_intercept_deepest_member |
1 |
0.0000 |
0.0000 |
1.0000 |
0 |
60 |
| catena_galatians |
no_intercept_deepest_member |
1 |
0.0000 |
0.0000 |
1.0000 |
0 |
60 |
| catena_romans |
no_intercept_deepest_member |
1 |
0.0167 |
0.0167 |
1.0000 |
0 |
60 |
| comp_tc_mark_stemma_backbone |
drop_zero_members |
1 |
0.2000 |
0.2000 |
11.5000 |
0 |
0 |
Member Diagnostics
| dataset |
subset_size |
include_intercept |
retained_member_rows |
retained_root_nearest_rate |
retained_intercept_nonzero_rate |
retained_prediction_equals_intercept_rate |
| catena_ephesians |
1 |
0 |
0 |
|
|
|
| catena_ephesians |
2 |
0 |
0 |
|
|
|
| catena_ephesians |
1 |
1 |
359 |
0.0000 |
1.0000 |
0.8162 |
| catena_galatians |
1 |
0 |
0 |
|
|
|
| catena_galatians |
2 |
0 |
0 |
|
|
|
| catena_galatians |
1 |
1 |
357 |
0.0000 |
1.0000 |
0.9216 |
| catena_romans |
1 |
0 |
0 |
|
|
|
| catena_romans |
2 |
0 |
0 |
|
|
|
| catena_romans |
1 |
1 |
358 |
0.0000 |
1.0000 |
0.8547 |
| comp_tc_mark_stemma_backbone |
1 |
0 |
60 |
0.0000 |
0.0000 |
0.0000 |
| comp_tc_mark_stemma_backbone |
2 |
0 |
136 |
0.0000 |
0.0000 |
0.0000 |
| comp_tc_mark_stemma_backbone |
1 |
1 |
359 |
0.0000 |
0.8468 |
0.8357 |
Tradeoff Diagnostics
| dataset |
family_points |
slope_log_loss_vs_log_params |
r_squared |
spearman_rho |
| catena_ephesians |
5 |
0.0606 |
0.3082 |
0.8947 |
| catena_galatians |
5 |
0.0467 |
0.3778 |
0.8947 |
| catena_romans |
5 |
0.0572 |
0.2893 |
0.8947 |
| comp_tc_mark_stemma_backbone |
5 |
-0.0629 |
0.3298 |
-0.5000 |
Skipped P-Adic Fits
| dataset |
panel_id |
subset_size |
include_intercept |
estimated_point_systems |
max_point_systems |
| catena_ephesians |
panel_coverage |
2 |
1 |
205320 |
100000 |
| catena_galatians |
panel_coverage |
2 |
1 |
205320 |
100000 |
| catena_romans |
panel_coverage |
2 |
1 |
205320 |
100000 |
| comp_tc_mark_stemma_backbone |
panel_coverage |
2 |
1 |
205320 |
100000 |
Outputs
- Member diagnostics:
outputs/tables/external_padic_aggregation_experiments_member_diagnostics.csv
- P-adic policy summary:
outputs/tables/external_padic_aggregation_experiments_padic_summary.csv
- Euclidean tie-scored summary:
outputs/tables/external_padic_aggregation_experiments_euclidean_summary.csv
- Combined comparison summary:
outputs/tables/external_padic_aggregation_experiments_summary.csv
- Story comparison:
outputs/tables/external_padic_aggregation_experiments_story.csv