Spatial vs random validation split — descriptive comparison

Spatial split source: /home/valerian/SGTPublication/Weights-ResidualsModels-MappingInference-SOCmapping/TemporalFusionTransformer/residualModels1mil_normalize_composite_l2_v2/train_val_data_run_1_MAX_OC_150_TIME_BEGINNING_2007_TIME_END_2023_TRANSFORM_normalize_LOSS_composite_l2.parquet

Random split: synthetic, drawn here with seed = 20260513. Same size (1359 rows) sampled without replacement from the full pool of 16514 rows in the same parquet.

Caveat. Every train/val parquet I found in the repository (residualModels1mil_normalize_*, _log_*, _normalize_mse) uses the same create_validation_train_sets(distance_threshold=1.2) spatial-split logic. There is no surviving parquet from a purely random split. The "random" column below is therefore a synthetic reference draw — it shows what a same-size random validation set looks like on this dataset, which is the closest you can get to an apples-to-apples random-vs-spatial comparison without retraining.

Side-by-side descriptive statistics

metric	spatial val	random val	Δ (random − spatial)
n	1359	1359	—
mean (g/kg)	16.337	22.903	+6.567
sd (g/kg)	7.474	20.290	+12.815
median (g/kg)	14.400	16.600	+2.200
q25 (g/kg)	12.400	12.625	+0.225
q75 (g/kg)	17.400	24.050	+6.650
IQR (g/kg)	5.000	11.425	+6.425
min (g/kg)	4.460	3.700	-0.760
max (g/kg)	85.000	147.000	+62.000
n > 50 g/kg	8	107	—
% > 50 g/kg	0.59%	7.87%	+7.285
n > 120 g/kg	0	12	—
% > 120 g/kg	0.00%	0.88%	+0.883

Geographic extent

split	lat range	lon range	unique locations
spatial_val	[47.3412, 50.4876]	[9.6002, 13.8337]	513
random_val	[47.2409, 50.5842]	[9.5407, 13.8768]	1088

Per-year sample count in each validation set

year	spatial val n	random val n
2007	32	116
2008	11	58
2009	94	63
2010	9	15
2011	246	176
2012	285	250
2013	192	220
2014	113	67
2015	203	113
2016	12	9
2017	19	111
2018	34	93
2019	12	16
2020	47	15
2021	40	19
2022	9	13
2023	1	5