Upload results/phase1_filtering_summary.txt with huggingface_hub
Browse files
results/phase1_filtering_summary.txt
ADDED
|
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
================================================================================
|
| 2 |
+
PHASE 1: ENVIRONMENTAL FILTERING METRICS
|
| 3 |
+
================================================================================
|
| 4 |
+
|
| 5 |
+
Provenance:
|
| 6 |
+
Script: /media/drn2/External/TARA-Oceans/03_analyses/WorldModelApp/scripts/t07_filtering_metrics_20260127_150000.py
|
| 7 |
+
Input: /media/drn2/External/TARA-Oceans/03_analyses/WorldModelApp/results/phase1_xgb_predictions.npz
|
| 8 |
+
Date: 2026-01-27 10:03:39
|
| 9 |
+
Integrity Check: PASSED
|
| 10 |
+
|
| 11 |
+
Dataset: 1810 samples, 20 PFAM modules
|
| 12 |
+
Method: Multi-output XGBoost (env -> PFAM modules), leave-one-basin-out CV
|
| 13 |
+
|
| 14 |
+
--------------------------------------------------------------------------------
|
| 15 |
+
COSINE SIMILARITY (predicted vs true PFAM module vectors)
|
| 16 |
+
--------------------------------------------------------------------------------
|
| 17 |
+
|
| 18 |
+
Mean: 0.8788
|
| 19 |
+
Median: 0.9295
|
| 20 |
+
Std: 0.1441
|
| 21 |
+
Min: -0.3170
|
| 22 |
+
Max: 0.9972
|
| 23 |
+
|
| 24 |
+
Percentile distribution:
|
| 25 |
+
P 5 = 0.5535
|
| 26 |
+
P10 = 0.6965
|
| 27 |
+
P25 = 0.8656
|
| 28 |
+
P50 = 0.9295
|
| 29 |
+
P75 = 0.9651
|
| 30 |
+
P90 = 0.9813
|
| 31 |
+
P95 = 0.9874
|
| 32 |
+
|
| 33 |
+
Fraction above threshold:
|
| 34 |
+
> 0.50: 96.0% (1737/1810)
|
| 35 |
+
> 0.80: 83.0% (1503/1810)
|
| 36 |
+
> 0.90: 63.7% (1153/1810)
|
| 37 |
+
> 0.95: 38.0% (687/1810)
|
| 38 |
+
|
| 39 |
+
--------------------------------------------------------------------------------
|
| 40 |
+
L2 ERROR (Euclidean distance)
|
| 41 |
+
--------------------------------------------------------------------------------
|
| 42 |
+
|
| 43 |
+
Mean: 8906.31
|
| 44 |
+
Median: 5871.31
|
| 45 |
+
Std: 8406.03
|
| 46 |
+
Min: 654.80
|
| 47 |
+
Max: 70019.22
|
| 48 |
+
|
| 49 |
+
--------------------------------------------------------------------------------
|
| 50 |
+
PER-BASIN COSINE SIMILARITY
|
| 51 |
+
--------------------------------------------------------------------------------
|
| 52 |
+
|
| 53 |
+
Basin n Mean Std Median Min Max
|
| 54 |
+
---------------------------------------------------------------
|
| 55 |
+
Arctic 38 0.9124 0.0589 0.9093 0.7719 0.9859
|
| 56 |
+
Atlantic 644 0.9084 0.1118 0.9438 0.2581 0.9972
|
| 57 |
+
Indian 195 0.8248 0.1981 0.9011 -0.3170 0.9968
|
| 58 |
+
Mediterranean 242 0.8293 0.1851 0.9061 0.2341 0.9953
|
| 59 |
+
Pacific 617 0.8859 0.1235 0.9340 0.2511 0.9971
|
| 60 |
+
Red_Sea 24 0.6949 0.2134 0.7219 0.1851 0.9809
|
| 61 |
+
Southern 50 0.9207 0.0695 0.9445 0.5978 0.9885
|
| 62 |
+
|
| 63 |
+
--------------------------------------------------------------------------------
|
| 64 |
+
PER-MODULE R2 (pooled held-out predictions)
|
| 65 |
+
--------------------------------------------------------------------------------
|
| 66 |
+
|
| 67 |
+
Module R2
|
| 68 |
+
------------------------
|
| 69 |
+
module_19 +0.0871
|
| 70 |
+
module_14 +0.0568
|
| 71 |
+
module_15 +0.0369
|
| 72 |
+
module_4 +0.0328
|
| 73 |
+
module_17 +0.0269
|
| 74 |
+
module_10 +0.0265
|
| 75 |
+
module_11 +0.0192
|
| 76 |
+
module_12 +0.0153
|
| 77 |
+
module_5 +0.0104
|
| 78 |
+
module_3 +0.0014
|
| 79 |
+
module_18 -0.0004
|
| 80 |
+
module_7 -0.0025
|
| 81 |
+
module_16 -0.0144
|
| 82 |
+
module_8 -0.0153
|
| 83 |
+
module_1 -0.0647
|
| 84 |
+
module_2 -0.0738
|
| 85 |
+
module_20 -0.1492
|
| 86 |
+
module_6 -0.1529
|
| 87 |
+
module_13 -0.1541
|
| 88 |
+
module_9 -0.1877
|
| 89 |
+
|
| 90 |
+
Mean R2: -0.0251
|
| 91 |
+
Median R2: +0.0005
|
| 92 |
+
Positive R2: 10/20
|
| 93 |
+
Negative R2: 10/20
|
| 94 |
+
|
| 95 |
+
--------------------------------------------------------------------------------
|
| 96 |
+
PER-FOLD MEAN R2
|
| 97 |
+
--------------------------------------------------------------------------------
|
| 98 |
+
|
| 99 |
+
Arctic (n= 38): mean R2 = -5.9435
|
| 100 |
+
Atlantic (n= 644): mean R2 = +0.0448
|
| 101 |
+
Indian (n= 219): mean R2 = -0.1898
|
| 102 |
+
Mediterranean (n= 242): mean R2 = +0.0996
|
| 103 |
+
Pacific (n= 617): mean R2 = -0.1439
|
| 104 |
+
Southern (n= 50): mean R2 = -1.0108
|
| 105 |
+
|
| 106 |
+
--------------------------------------------------------------------------------
|
| 107 |
+
INTERPRETATION
|
| 108 |
+
--------------------------------------------------------------------------------
|
| 109 |
+
|
| 110 |
+
Environmental filtering strength is quantified by two metrics:
|
| 111 |
+
|
| 112 |
+
1. COSINE SIMILARITY (directional match): mean=0.8788
|
| 113 |
+
- High cosine similarity indicates the environment captures the DIRECTION
|
| 114 |
+
of PFAM module composition (which modules are relatively up vs down)
|
| 115 |
+
- 1503/1810 (83.0%) samples have cosine sim > 0.80
|
| 116 |
+
|
| 117 |
+
2. R2 PER MODULE (magnitude prediction): mean=-0.0251
|
| 118 |
+
- 10/20 modules have positive R2 (pooled)
|
| 119 |
+
- R2 near zero indicates the environment captures marginal
|
| 120 |
+
signal about absolute PFAM module abundances
|
| 121 |
+
|
| 122 |
+
NOTE: Cosine similarity may be inflated by dominant modules with
|
| 123 |
+
universally high counts. Permutation test (T08) needed to assess
|
| 124 |
+
whether observed cosine similarity exceeds null expectation.
|