Add support for SAX and PAA for CF sparsity validation (#65)

src/wildboar/metrics/_counterfactual.py

@@ -7,6 +7,10 @@
 
 from ..distance import paired_distance, pairwise_distance
 from ..explain.counterfactual import proximity
+from ..transform import (
+    piecewice_aggregate_approximation,
+    symbolic_aggregate_approximation,
+)
 from ..utils.validation import check_array
 
 
@@ -276,7 +280,15 @@ def proximity_score(
     )
 
 
-def compactness_score(x_true, x_counterfactuals, *, rtol=1.0e-5, atol=1.0e-8):
+def compactness_score(
+    x_true,
+    x_counterfactuals,
+    *,
+    window=None,
+    n_bins=None,
+    atol=1.0e-8,
+    average=True,
+):
     """Return the compactness of the counterfactuals as measured by the
     fraction of changed timesteps. The fewer timesteps have changed between the original
     and the counterfactual, the lower the score.
@@ -291,11 +303,14 @@ def compactness_score(x_true, x_counterfactuals, *, rtol=1.0e-5, atol=1.0e-8):
     or (n_samples, n_dims, n_timeteps)
         The counterfactual samples
 
-    rtol : float, optional
-        Parameter to `np.isclose`.
+    atol : float,
+        The absolute tolerance.
+
+    window : int, optional
+        If set, evaluate the difference between windows of specified size.
 
-    atol : float, optional
-        Parameter to `np.isclose`.
+    n_bins : int, optional
+        If set, evaluate the set overlap of SAX transformed series.
 
     Returns
     -------
@@ -316,7 +331,33 @@ def compactness_score(x_true, x_counterfactuals, *, rtol=1.0e-5, atol=1.0e-8):
             % (x_true.shape, x_counterfactuals.shape)
         )
 
-    return 1 - np.mean(np.isclose(x_counterfactuals, x_true, rtol=rtol, atol=atol))
+    if window is not None:
+        if n_bins is not None:
+
+            def score(x_counterfactuals, x_true):
+                x_counterfactuals = symbolic_aggregate_approximation(
+                    x_counterfactuals, window=window, n_bins=n_bins
+                )
+                x_true = symbolic_aggregate_approximation(
+                    x_true, window=window, n_bins=n_bins
+                )
+                return x_counterfactuals == x_true
+
+        else:
+
+            def score(x_counterfactuals, x_true):
+                x_counterfactuals = piecewice_aggregate_approximation(
+                    x_counterfactuals, window=window
+                )
+                x_true = piecewice_aggregate_approximation(x_true, window=window)
+                return np.mean(np.isclose(x_counterfactuals, x_true, rtol=0, atol=atol))
+
+    else:
+
+        def score(x_counterfactuals, x_true):
+            return np.isclose(x_counterfactuals, x_true, rtol=0, atol=atol)
+
+    return 1 - np.mean(score(x_counterfactuals, x_true), axis=None if average else 0)
 
 
 def validity_score(y_pred, y_counterfactual, sample_weight=None):