[ENH] Added ROCKAD anomaly detector to aeon (#2376)

* Added ROCKAD anomaly detector to aeon

* Added ROCKAD to anomaly_detection.rst

* Empty commit for CI

* Automatic `pre-commit` fixes

* Fix newline at end of file

* adopted code to fit refactored Rocket arguments

* added "capability:multithreading": True to _tags

* Catch power transform and disable it if it results in error

* Added fallback when power transform fails, added ValueError if number of windows is smaller than n.neighors and other general improvements

* Added private attribute for power_transform activation/deactivation

* added tests for kneighbors check, adapted univariate and multivariate tests for changes in ROCKAD

* Removed pandas, set rocket normalise default to False, use transposed data for power transform to prevent bracket error,  set cleaned up code. Set standardize=True on power transform so that StandardScaler can be removed.

* Added  test for power transform failure

* removed transpose, added back user warning when power transform fails, added comments for clarity, other small improvements

* removed noop

* removed inf_columns_ check

* moved parent class init and n_jobs to have consistent structure

---------

Co-authored-by: MatthewMiddlehurst <[email protected]>
Co-authored-by: pattplatt <[email protected]>

aeon/anomaly_detection/__init__.py

@@ -10,6 +10,7 @@
     "LOF",
     "MERLIN",
     "OneClassSVM",
+    "ROCKAD",
     "PyODAdapter",
     "STOMP",
     "STRAY",
@@ -25,5 +26,6 @@
 from aeon.anomaly_detection._merlin import MERLIN
 from aeon.anomaly_detection._one_class_svm import OneClassSVM
 from aeon.anomaly_detection._pyodadapter import PyODAdapter
+from aeon.anomaly_detection._rockad import ROCKAD
 from aeon.anomaly_detection._stomp import STOMP
 from aeon.anomaly_detection._stray import STRAY

aeon/anomaly_detection/_rockad.py

@@ -0,0 +1,262 @@
+"""ROCKAD anomaly detector."""
+
+__all__ = ["ROCKAD"]
+
+import warnings
+from typing import Optional
+
+import numpy as np
+from sklearn.neighbors import NearestNeighbors
+from sklearn.preprocessing import PowerTransformer
+from sklearn.utils import resample
+
+from aeon.anomaly_detection.base import BaseAnomalyDetector
+from aeon.transformations.collection.convolution_based import Rocket
+from aeon.utils.windowing import reverse_windowing, sliding_windows
+
+
+class ROCKAD(BaseAnomalyDetector):
+    """
+    ROCKET-based Anomaly Detector (ROCKAD).
+
+    ROCKAD leverages the ROCKET transformation for feature extraction from
+    time series data and applies the scikit learn k-nearest neighbors (k-NN)
+    approach with bootstrap aggregation for robust anomaly detection.
+    After windowing, the data gets transformed into the ROCKET feature space.
+    Then the windows are compared based on the feature space by
+    finding the nearest neighbours.
+
+    This class supports both univariate and multivariate time series and
+    provides options for normalizing features, applying power transformations,
+    and customizing the distance metric.
+
+    Parameters
+    ----------
+    n_estimators : int, default=10
+        Number of k-NN estimators to use in the bootstrap aggregation.
+    n_kernels : int, default=100
+        Number of kernels to use in the ROCKET transformation.
+    normalise : bool, default=False
+        Whether to normalize the ROCKET-transformed features.
+    n_neighbors : int, default=5
+        Number of neighbors to use for the k-NN algorithm.
+    n_jobs : int, default=1
+        Number of parallel jobs to use for the k-NN algorithm and ROCKET transformation.
+    metric : str, default="euclidean"
+        Distance metric to use for the k-NN algorithm.
+    power_transform : bool, default=True
+        Whether to apply a power transformation (Yeo-Johnson) to the features.
+    window_size : int, default=10
+        Size of the sliding window for segmenting input time series data.
+    stride : int, default=1
+        Step size for moving the sliding window over the time series data.
+    random_state : int, default=42
+        Random seed for reproducibility.
+
+    Attributes
+    ----------
+    rocket_transformer_ : Optional[Rocket]
+        Instance of the ROCKET transformer used to extract features, set after fitting.
+    list_baggers_ : Optional[list[NearestNeighbors]]
+        List containing k-NN estimators used for anomaly scoring, set after fitting.
+    power_transformer_ : PowerTransformer
+        Transformer used to apply power transformation to the features.
+    """
+
+    _tags = {
+        "capability:univariate": True,
+        "capability:multivariate": True,
+        "capability:missing_values": False,
+        "capability:multithreading": True,
+        "fit_is_empty": False,
+    }
+
+    def __init__(
+        self,
+        n_estimators=10,
+        n_kernels=100,
+        normalise=False,
+        n_neighbors=5,
+        metric="euclidean",
+        power_transform=True,
+        window_size: int = 10,
+        stride: int = 1,
+        n_jobs=1,
+        random_state=42,
+    ):
+
+        self.n_estimators = n_estimators
+        self.n_kernels = n_kernels
+        self.normalise = normalise
+        self.n_neighbors = n_neighbors
+        self.n_jobs = n_jobs
+        self.metric = metric
+        self.power_transform = power_transform
+        self.window_size = window_size
+        self.stride = stride
+        self.random_state = random_state
+
+        self.rocket_transformer_: Optional[Rocket] = None
+        self.list_baggers_: Optional[list[NearestNeighbors]] = None
+        self.power_transformer_: Optional[PowerTransformer] = None
+
+        super().__init__(axis=0)
+
+    def _fit(self, X: np.ndarray, y: Optional[np.ndarray] = None) -> "ROCKAD":
+        self._check_params(X)
+        # X: (n_timepoints, 1) because __init__(axis==0)
+        _X, _ = sliding_windows(
+            X, window_size=self.window_size, stride=self.stride, axis=0
+        )
+        # _X: (n_windows, window_size)
+        self._inner_fit(_X)
+
+        return self
+
+    def _check_params(self, X: np.ndarray) -> None:
+        if self.window_size < 1 or self.window_size > X.shape[0]:
+            raise ValueError(
+                "The window size must be at least 1 and at most the length of the "
+                "time series."
+            )
+
+        if self.stride < 1 or self.stride > self.window_size:
+            raise ValueError(
+                "The stride must be at least 1 and at most the window size."
+            )
+
+        if int((X.shape[0] - self.window_size) / self.stride + 1) < self.n_neighbors:
+            raise ValueError(
+                f"Window count ({int((X.shape[0]-self.window_size)/self.stride+1)}) "
+                f"has to be larger than n_neighbors ({self.n_neighbors})."
+                "Please choose a smaller n_neighbors value or increase window count "
+                "by choosing a smaller window size or larger stride."
+            )
+
+    def _inner_fit(self, X: np.ndarray) -> None:
+
+        self.rocket_transformer_ = Rocket(
+            n_kernels=self.n_kernels,
+            normalise=self.normalise,
+            n_jobs=self.n_jobs,
+            random_state=self.random_state,
+        )
+        # X: (n_windows, window_size)
+        Xt = self.rocket_transformer_.fit_transform(X)
+        # XT: (n_cases, n_kernels*2)
+        Xt = Xt.astype(np.float64)
+
+        if self.power_transform:
+            self.power_transformer_ = PowerTransformer()
+            try:
+                Xtp = self.power_transformer_.fit_transform(Xt)
+
+            except Exception:
+                warnings.warn(
+                    "Power Transform failed and thus has been disabled. "
+                    "Try increasing the window size.",
+                    UserWarning,
+                    stacklevel=2,
+                )
+                self.power_transformer_ = None
+                Xtp = Xt
+        else:
+            Xtp = Xt
+
+        self.list_baggers_ = []
+
+        for idx_estimator in range(self.n_estimators):
+            # Initialize estimator
+            estimator = NearestNeighbors(
+                n_neighbors=self.n_neighbors,
+                n_jobs=self.n_jobs,
+                metric=self.metric,
+                algorithm="kd_tree",
+            )
+            # Bootstrap Aggregation
+            Xtp_scaled_sample = resample(
+                Xtp,
+                replace=True,
+                n_samples=None,
+                random_state=self.random_state + idx_estimator,
+                stratify=None,
+            )
+
+            # Fit estimator and append to estimator list
+            estimator.fit(Xtp_scaled_sample)
+            self.list_baggers_.append(estimator)
+
+    def _predict(self, X) -> np.ndarray:
+
+        _X, padding = sliding_windows(
+            X, window_size=self.window_size, stride=self.stride, axis=0
+        )
+
+        point_anomaly_scores = self._inner_predict(_X, padding)
+
+        return point_anomaly_scores
+
+    def _fit_predict(self, X: np.ndarray, y: Optional[np.ndarray] = None) -> np.ndarray:
+        self._check_params(X)
+        _X, padding = sliding_windows(
+            X, window_size=self.window_size, stride=self.stride, axis=0
+        )
+
+        self._inner_fit(_X)
+        point_anomaly_scores = self._inner_predict(_X, padding)
+        return point_anomaly_scores
+
+    def _inner_predict(self, X: np.ndarray, padding: int) -> np.ndarray:
+
+        anomaly_scores = self._predict_proba(X)
+
+        point_anomaly_scores = reverse_windowing(
+            anomaly_scores, self.window_size, np.nanmean, self.stride, padding
+        )
+
+        point_anomaly_scores = (point_anomaly_scores - point_anomaly_scores.min()) / (
+            point_anomaly_scores.max() - point_anomaly_scores.min()
+        )
+
+        return point_anomaly_scores
+
+    def _predict_proba(self, X):
+        """
+        Predicts the probability of anomalies for the input data.
+
+        Parameters
+        ----------
+            X (array-like): The input data.
+
+        Returns
+        -------
+            np.ndarray: The predicted probabilities.
+
+        """
+        y_scores = np.zeros((len(X), self.n_estimators))
+        # Transform into rocket feature space
+        Xt = self.rocket_transformer_.transform(X)
+
+        Xt = Xt.astype(np.float64)
+
+        if self.power_transformer_ is not None:
+            # Power Transform using yeo-johnson
+            Xtp = self.power_transformer_.transform(Xt)
+
+        else:
+            Xtp = Xt
+
+        for idx, bagger in enumerate(self.list_baggers_):
+            # Get scores from each estimator
+            distances, _ = bagger.kneighbors(Xtp)
+
+            # Compute mean distance of nearest points in window
+            scores = distances.mean(axis=1).reshape(-1, 1)
+            scores = scores.squeeze()
+
+            y_scores[:, idx] = scores
+
+        # Average the scores to get the final score for each time series
+        y_scores = y_scores.mean(axis=1)
+
+        return y_scores

aeon/anomaly_detection/tests/test_rockad.py

@@ -0,0 +1,75 @@
+"""Tests for the ROCKAD anomaly detector."""
+
+import numpy as np
+import pytest
+from sklearn.utils import check_random_state
+
+from aeon.anomaly_detection import ROCKAD
+
+
+def test_rockad_univariate():
+    """Test ROCKAD univariate output."""
+    rng = check_random_state(seed=2)
+    series = rng.normal(size=(100,))
+    series[50:58] -= 5
+
+    ad = ROCKAD(
+        n_estimators=100,
+        n_kernels=10,
+        n_neighbors=9,
+        power_transform=True,
+        window_size=20,
+        stride=1,
+    )
+
+    pred = ad.fit_predict(series, axis=0)
+
+    assert pred.shape == (100,)
+    assert pred.dtype == np.float64
+    assert 50 <= np.argmax(pred) <= 58
+
+
+def test_rockad_multivariate():
+    """Test ROCKAD multivariate output."""
+    rng = check_random_state(seed=2)
+    series = rng.normal(size=(100, 3))
+    series[50:58, 0] -= 5
+    series[87:90, 1] += 0.1
+
+    ad = ROCKAD(
+        n_estimators=1000,
+        n_kernels=100,
+        n_neighbors=20,
+        power_transform=True,
+        window_size=10,
+        stride=1,
+    )
+
+    pred = ad.fit_predict(series, axis=0)
+
+    assert pred.shape == (100,)
+    assert pred.dtype == np.float64
+    assert 50 <= np.argmax(pred) <= 58
+
+
+def test_rockad_incorrect_input():
+    """Test ROCKAD incorrect input."""
+    rng = check_random_state(seed=2)
+    series = rng.normal(size=(100,))
+
+    with pytest.raises(ValueError, match="The window size must be at least 1"):
+        ad = ROCKAD(window_size=0)
+        ad.fit_predict(series)
+    with pytest.raises(ValueError, match="The stride must be at least 1"):
+        ad = ROCKAD(stride=0)
+        ad.fit_predict(series)
+    with pytest.raises(
+        ValueError, match=r"Window count .* has to be larger than n_neighbors .*"
+    ):
+        ad = ROCKAD(stride=1, window_size=100)
+        ad.fit_predict(series)
+    with pytest.warns(
+        UserWarning, match=r"Power Transform failed and thus has been disabled."
+    ):
+        ad = ROCKAD(stride=1, window_size=5)
+        ad.fit_predict(series)

docs/api_reference/anomaly_detection.rst

@@ -32,6 +32,7 @@ Detectors
     MERLIN
     OneClassSVM
     PyODAdapter
+    ROCKAD
     STOMP
     STRAY