From 5aaa4baa176ba710c219b17529eb67cdc2cad257 Mon Sep 17 00:00:00 2001
From: Ivan Knyazev <105492484+IRKnyazev@users.noreply.github.com>
Date: Wed, 28 Aug 2024 14:19:19 +0700
Subject: [PATCH] [BUG] Short term solution for shapelet feature importance
 (#2017)

* partial fixed to issue

* added a warning message for rdst combined with linear classifiers
---
 aeon/visualisation/estimator/_shapelets.py | 31 +++++++++++++++++++---
 1 file changed, 27 insertions(+), 4 deletions(-)

diff --git a/aeon/visualisation/estimator/_shapelets.py b/aeon/visualisation/estimator/_shapelets.py
index 847a96381d..1feeb3ee04 100644
--- a/aeon/visualisation/estimator/_shapelets.py
+++ b/aeon/visualisation/estimator/_shapelets.py
@@ -5,6 +5,7 @@
 __all__ = ["ShapeletClassifierVisualizer", "ShapeletTransformerVisualizer"]
 
 import copy
+import warnings
 
 import numpy as np
 from sklearn.ensemble._forest import BaseForest
@@ -681,13 +682,35 @@ def _get_shp_importance(self, class_id):
         if isinstance(classifier, Pipeline):
             classifier = classifier[-1]
 
-        # This suppose that the higher the coef linked to each feature, the most
-        # impact this feature makes on classification for the given class_id
+        # This supposes that the higher (with the exception of distance features)
+        # the coef linked to each feature, the most impact this feature makes on
+        # classification for the given class_id
         if isinstance(classifier, LinearClassifierMixin):
             coefs = classifier.coef_
             n_classes = coefs.shape[0]
             if n_classes == 1:
-                coefs = np.append(-coefs, coefs, axis=0)
+                if isinstance(self.estimator, RDSTClassifier):
+                    class_0_coefs = np.copy(coefs)
+                    class_1_coefs = np.copy(coefs)
+
+                    mask = np.ones(class_0_coefs.shape[1], dtype=bool)
+                    mask[::3] = False
+                    class_0_coefs[:, mask] = -class_0_coefs[:, mask]
+                    class_1_coefs[:, ::3] = -class_1_coefs[:, ::3]
+
+                    # Append the two modified coefs arrays along axis 0
+                    coefs = np.append(class_0_coefs, class_1_coefs, axis=0)
+                    warnings.warn(
+                        "Shapelet importance ranking may be unreliable "
+                        "when using linear classifiers with RDST. "
+                        "This is due to the interaction between argmin "
+                        "and shapelet occurrence features, which can distort "
+                        "the rankings. Consider evaluating the results carefully "
+                        "or using an alternative method.",
+                        stacklevel=1,
+                    )
+                else:
+                    coefs = np.append(coefs, -coefs, axis=0)
             coefs = coefs[class_id]
 
         elif isinstance(classifier, (BaseForest, BaseDecisionTree)):
@@ -699,7 +722,7 @@ def _get_shp_importance(self, class_id):
                 "classifier inheriting from LinearClassifierMixin, BaseForest or "
                 f"BaseDecisionTree but got {type(classifier)}"
             )
-        # coefs = coefs[idx]
+
         if isinstance(self.estimator, RDSTClassifier):
             # As each shapelet generate 3 features, divide feature id by 3 so all
             # features generated by one shapelet share the same ID