Rewritten implementations of the KLIEP and RuLSIF algorithms #57
Conversation
| :param detected_change_points: list of detected change point indices. | ||
| :return: a dictionary with evaluation metrics (precision, recall, F1 score). | ||
| def evaluate_detection_accuracy( |
There was a problem hiding this comment.
Do we really need this here? It seems this logic belongs to benchmarking.
| @@ -1,112 +1,155 @@ | |||
| from abc import abstractmethod | |||
There was a problem hiding this comment.
Add license/copyright metadata here (and header docstring)
| :param window: input data window | ||
| :return: True if window can be processed, else False | ||
| """ | ||
| return len(window) >= 2 * self.min_window_size |
There was a problem hiding this comment.
Why len must be at least twice as the minimal? It looks a little bit confusing.
| def loss(alpha_array: npt.NDArray[np.float64], /) -> float: | ||
| alpha = alpha_array[0] | ||
| ratio = np.exp(np.log(test_density) - np.log(ref_density + 1e-10) - alpha) | ||
| return float(-np.mean(np.log(ratio + 1e-10)) + self.regularization * alpha**2) |
There was a problem hiding this comment.
What are these 1e-10? Is it for numerical stability? Should it be a parameter? Or maybe make it a private field of class?
| ) | ||
|
|
||
| return cast(list[int], np.where(weights > self.threshold)[0].tolist()) | ||
| def _kde_on_grid( |
There was a problem hiding this comment.
Method '_kde_on_grid' may be 'static'
|
|
||
| def objective_function_wrapper(alpha: npt.NDArray[np.float64], /) -> float: | ||
| """Wrapper for the objective function to calculate the density ratio. | ||
| def _validate_window(self, window: npt.NDArray[Any]) -> npt.NDArray[np.float64]: |
There was a problem hiding this comment.
Method '_validate_window' may be 'static'
| test_density = self._kernel_density_estimation(test_value, bandwidth) | ||
| reference_density = self._kernel_density_estimation(reference_value, bandwidth) | ||
| window = self._validate_window(window) | ||
| if not self._is_window_valid(window): |
There was a problem hiding this comment.
Why previously validated window can be not valid here? Naming looks a little bit confusing.
Based on #43