[ENH] homogenization of sktime and skchange detection API - predict…

…, `predict_points`, `predict_segments` (#7476) This PR carries out some remaining steps of interface homogenization with `skchange`: * index of returns of `predict`, `predict_points`, `predict_segments` is now always a `RangeIndex` - no key information requires index manipulation anymore. * return of `predict`, `predict_points`, `predict_segments` is now always a `pd.DataFrame`, with one column `"ilocs"` for the event indices, and additional potential columns in supervised and semi-supervised cases * return of `predict` is now identical with one of `predict_points` or `predict_segments` in the supervised and semi-supervised case Also updates some estimators to function as expected: * existing estimators through coercions to the desired output * dummy estimators
sktime · Dec 3, 2024 · 9bf1324 · 9bf1324
1 parent 1d84df2
commit 9bf1324
Show file tree

Hide file tree

Showing 15 changed files with 326 additions and 80 deletions.
diff --git a/examples/07_detection_anomaly_changepoints.ipynb b/examples/07_detection_anomaly_changepoints.ipynb
@@ -870,7 +870,7 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 17,
+   "execution_count": null,
    "id": "6fcf2dbd-0afb-4d11-b238-84294e811dba",
    "metadata": {},
    "outputs": [
@@ -905,7 +905,7 @@
     "    linestyle=\"--\",\n",
     ")\n",
     "\n",
-    "for i, cp in enumerate(predicted_change_points):\n",
+    "for i, cp in enumerate(predicted_change_points.values.flatten()):\n",
     "    label = \"Predicted Change Points\" if i == 0 else None\n",
     "    ax.axvline(cp, color=\"tab:green\", linestyle=\"--\", label=label)\n",
     "\n",

diff --git a/extension_templates/detection.py b/extension_templates/detection.py
@@ -192,8 +192,29 @@ def _fit(self, X, y=None):
         ----------
         X : pd.DataFrame
             Training data to fit model to time series.
-        y : pd.Series, optional
-            Ground truth labels for training, if detector is supervised.
+
+        y : pd.DataFrame with RangeIndex
+            Known events for training, in ``X``, if detector is supervised.
+
+            Each row ``y`` is a known event.
+            Can have the following columns:
+
+            * ``"ilocs"`` - always. Values encode where/when the event takes place,
+              via ``iloc`` references to indices of ``X``,
+              or ranges ot indices of ``X``, as below.
+            * ``"label"`` - if the task, by tags, is supervised or semi-supervised
+              segmentation with labels, or segment clustering.
+
+            The meaning of entries in the ``"ilocs"`` column and ``"labels"``
+            column describe the event in a given row as follows:
+
+            * If ``task`` is ``"anomaly_detection"`` or ``"change_point_detection"``,
+              ``"ilocs"`` contains the iloc index at which the event takes place.
+            * If ``task`` is ``"segmentation"``, ``"ilocs"`` contains left-closed
+              intervals of iloc based segments, interpreted as the range
+              of indices over which the event takes place.
+
+            Labels (if present) in the ``"labels"`` column indicate the type of event.
 
         Returns
         -------
@@ -221,13 +242,28 @@ def _predict(self, X):
 
         Returns
         -------
-        y : pd.Series with RangeIndex
-            Labels for sequence ``X``, in sparse format.
-            Values are ``iloc`` references to indices of ``X``.
+        y : pd.DataFrame with RangeIndex
+            Detected or predicted events.
+
+            Each row ``y`` is a detected or predicted event.
+            Can have the following columns:
+
+            * ``"ilocs"`` - always. Values encode where/when the event takes place,
+              via ``iloc`` references to indices of ``X``,
+              or ranges ot indices of ``X``, as below.
+            * ``"label"`` - if the task, by tags, is supervised or semi-supervised
+              segmentation with labels, or segment clustering.
+
+            The meaning of entries in the ``"ilocs"`` column and ``"labels"``
+            column describe the event in a given row as follows:
 
             * If ``task`` is ``"anomaly_detection"`` or ``"change_point_detection"``,
-              the values are integer indices of the changepoints/anomalies.
-            * If ``task`` is "segmentation", the values are ``pd.Interval`` objects.
+              ``"ilocs"`` contains the iloc index at which the event takes place.
+            * If ``task`` is ``"segmentation"``, ``"ilocs"`` contains left-closed
+              intervals of iloc based segments, interpreted as the range
+              of indices over which the event takes place.
+
+            Labels (if present) in the ``"labels"`` column indicate the type of event.
         """
 
         # implement here

diff --git a/sktime/detection/adapters/_pyod.py b/sktime/detection/adapters/_pyod.py
@@ -102,7 +102,7 @@ def _predict(self, X):
             Y_val_np = Y_np
 
         Y_loc = np.where(Y_np)
-        Y = pd.Series(Y_val_np[Y_loc], index=X.index[Y_loc])
+        Y = pd.Series(Y_val_np[Y_loc])
 
         return Y