implementation of spearman correlation

Author: ahinterl94-th

src/spotpython/light/trainmodel.py

@@ -1,7 +1,7 @@
 import lightning as L
 from spotpython.data.lightdatamodule import LightDataModule, PadSequenceManyToMany
 from spotpython.utils.eda import generate_config_id
-from spotpython.utils.metrics import calculate_xai_consistency_corr, calculate_xai_consistency_cosine, calculate_xai_consistency_euclidean
+from spotpython.utils.metrics import calculate_xai_consistency_corr, calculate_xai_consistency_cosine, calculate_xai_consistency_euclidean, calculate_xai_consistency_spearman
 from pytorch_lightning.loggers import TensorBoardLogger
 from lightning.pytorch.callbacks.early_stopping import EarlyStopping
 from lightning.pytorch.callbacks import ModelCheckpoint
@@ -723,13 +723,15 @@ def train_model_xai(config: dict, fun_control: dict, timestamp: bool = True) ->
     attributions = np.stack(attributions_list, axis=0)
 
     # Calculate corr:
-    if fun_control["xai_metric"] not in ["corr", "cosine", "euclidean"]:
-        raise ValueError(f"Invalid xai_metric: {fun_control['xai_metric']}. Valid metrics are: 'corr', 'cosine', 'euclidean'")
+    if fun_control["xai_metric"] not in ["corr", "cosine", "euclidean", "spearman"]:
+        raise ValueError(f"Invalid xai_metric: {fun_control['xai_metric']}. Valid metrics are: 'corr', 'cosine', 'euclidean', 'spearman'")
     if fun_control["xai_metric"] == "corr":
         result_xai = calculate_xai_consistency_corr(attributions)
     elif fun_control["xai_metric"] == "cosine":
         result_xai = calculate_xai_consistency_cosine(attributions)
     elif fun_control["xai_metric"] == "euclidean":
         result_xai = calculate_xai_consistency_euclidean(attributions)
+    elif fun_control["xai_metric"] == "spearman":
+        result_xai = calculate_xai_consistency_spearman(attributions)
 
     return result["val_loss"], result_xai

src/spotpython/utils/metrics.py

@@ -27,6 +27,8 @@
 import numpy as np
 from spotpython.utils.convert import series_to_array
 from sklearn.metrics.pairwise import euclidean_distances
+from scipy.stats import spearmanr
+
 
 
 def apk(actual, predicted, k=10):
@@ -275,3 +277,32 @@ def calculate_xai_consistency_euclidean(attributions):
     print("XAI Consistency (mean of upper triangle of Euclidean distance matrix):")
     print(result_xai)
     return result_xai
+
+def calculate_xai_consistency_spearman(attributions):
+    """
+    Calculates the consistency of XAI methods using Spearman rank correlation.
+    
+    Args:
+        attributions (np.ndarray): shape (n_methods, n_features)
+    
+    Returns:
+        float: Mean of upper triangle of Spearman correlation matrix (excluding diagonal)
+    """
+    attributions = np.array(attributions)
+    n_methods = attributions.shape[0]
+
+    spearman_corr_matrix = np.zeros((n_methods, n_methods))
+    for i in range(n_methods):
+        for j in range(n_methods):
+            corr = spearmanr(attributions[i], attributions[j]).correlation
+            if np.isnan(corr):
+                corr = 0.0
+            spearman_corr_matrix[i, j] = corr
+
+    upper_triangle_values = spearman_corr_matrix[np.triu_indices(n_methods, k=1)]
+    print("Attribution Spearman Correlation Matrix:")
+    print(spearman_corr_matrix)
+
+    print("XAI Consistency (mean of upper triangle of Spearman correlation matrix):")
+    print(upper_triangle_values.mean())
+    return upper_triangle_values.mean()