sliceline 0.2.18__py3-none-any.whl → 0.3.0__py3-none-any.whl

sliceline/__init__.py

@@ -1,3 +1,3 @@
-from .slicefinder import Slicefinder
+from .slicefinder import Slicefinder, is_numba_available
 
-__all__ = ("Slicefinder",)
+__all__ = ("Slicefinder", "is_numba_available")

sliceline/_numba_ops.py

@@ -0,0 +1,245 @@
+"""Numba-accelerated operations for Sliceline.
+
+Provides JIT-compiled versions of performance-critical functions
+for 5-50x performance improvements in scoring and ID computation.
+
+This module is optional - if numba is not installed, the main slicefinder
+module will fall back to pure NumPy implementations.
+
+Installation:
+    pip install numba
+    # or
+    pip install sliceline[optimized]
+"""
+
+from __future__ import annotations
+
+import numpy as np
+from numba import njit
+
+
+@njit(cache=True)
+def score_slices_numba(
+    slice_sizes: np.ndarray,
+    slice_errors: np.ndarray,
+    n_row: int,
+    alpha: float,
+    avg_error: float,
+) -> np.ndarray:
+    """JIT-compiled slice scoring function.
+
+    Computes scores for each slice based on size and error metrics.
+    5-10x faster than pure NumPy implementation.
+
+    Parameters
+    ----------
+    slice_sizes : np.ndarray
+        Array of slice sizes.
+    slice_errors : np.ndarray
+        Array of slice errors.
+    n_row : int
+        Number of rows in the encoded dataset.
+    alpha : float
+        Weight parameter for error importance.
+    avg_error : float
+        Average error across all samples.
+
+    Returns
+    -------
+    np.ndarray
+        Array of computed scores for each slice.
+    """
+    n = slice_sizes.shape[0]
+    scores = np.empty(n, dtype=np.float64)
+
+    for i in range(n):
+        if slice_sizes[i] <= 0:
+            scores[i] = -np.inf
+        else:
+            slice_avg_error = slice_errors[i] / slice_sizes[i]
+            error_term = alpha * (slice_avg_error / avg_error - 1.0)
+            size_term = (1.0 - alpha) * (n_row / slice_sizes[i] - 1.0)
+            scores[i] = error_term - size_term
+
+    return scores
+
+
+@njit(cache=True)
+def score_ub_single_numba(
+    slice_size: float,
+    slice_error: float,
+    max_slice_error: float,
+    n_col_x_encoded: int,
+    min_sup: float,
+    alpha: float,
+    avg_error: float,
+) -> float:
+    """JIT-compiled upper bound score for a single slice.
+
+    Parameters
+    ----------
+    slice_size : float
+        Size of the slice.
+    slice_error : float
+        Error sum of the slice.
+    max_slice_error : float
+        Maximum error in the slice.
+    n_col_x_encoded : int
+        Number of encoded columns.
+    min_sup : float
+        Minimum support threshold.
+    alpha : float
+        Weight parameter for error importance.
+    avg_error : float
+        Average error across all samples.
+
+    Returns
+    -------
+    float
+        Upper bound score for the slice.
+    """
+    if slice_size <= 0:
+        return -np.inf
+
+    potential_solutions = np.array(
+        [
+            min_sup,
+            max(slice_error / max_slice_error, min_sup)
+            if max_slice_error > 0
+            else min_sup,
+            slice_size,
+        ]
+    )
+
+    max_score = -np.inf
+    for pot_sol in potential_solutions:
+        if pot_sol <= 0:
+            continue
+        error_contrib = min(pot_sol * max_slice_error, slice_error)
+        score = (
+            alpha * (error_contrib / avg_error - pot_sol)
+            - (1.0 - alpha) * (n_col_x_encoded - pot_sol)
+        ) / pot_sol
+        if score > max_score:
+            max_score = score
+
+    return max_score
+
+
+@njit(cache=True)
+def score_ub_batch_numba(
+    slice_sizes_ub: np.ndarray,
+    slice_errors_ub: np.ndarray,
+    max_slice_errors_ub: np.ndarray,
+    n_col_x_encoded: int,
+    min_sup: float,
+    alpha: float,
+    avg_error: float,
+) -> np.ndarray:
+    """JIT-compiled upper bound scoring function for batch processing.
+
+    5-10x faster than pure NumPy implementation.
+
+    Parameters
+    ----------
+    slice_sizes_ub : np.ndarray
+        Array of slice sizes (upper bound).
+    slice_errors_ub : np.ndarray
+        Array of slice errors (upper bound).
+    max_slice_errors_ub : np.ndarray
+        Array of maximum slice errors (upper bound).
+    n_col_x_encoded : int
+        Number of encoded columns.
+    min_sup : float
+        Minimum support threshold.
+    alpha : float
+        Weight parameter for error importance.
+    avg_error : float
+        Average error across all samples.
+
+    Returns
+    -------
+    np.ndarray
+        Array of upper bound scores for each slice.
+    """
+    n = slice_sizes_ub.shape[0]
+    scores = np.empty(n, dtype=np.float64)
+
+    for i in range(n):
+        scores[i] = score_ub_single_numba(
+            slice_sizes_ub[i],
+            slice_errors_ub[i],
+            max_slice_errors_ub[i],
+            n_col_x_encoded,
+            min_sup,
+            alpha,
+            avg_error,
+        )
+
+    return scores
+
+
+@njit(cache=True)
+def compute_slice_ids_numba(
+    slices_data: np.ndarray,
+    slices_indices: np.ndarray,
+    slices_indptr: np.ndarray,
+    feature_offset_start: np.ndarray,
+    feature_offset_end: np.ndarray,
+    feature_domains: np.ndarray,
+) -> np.ndarray:
+    """JIT-compiled slice ID computation.
+
+    Computes unique IDs for each slice based on feature encoding.
+    10-50x faster than Python loop for large datasets.
+
+    Parameters
+    ----------
+    slices_data : np.ndarray
+        Data array from sparse matrix.
+    slices_indices : np.ndarray
+        Column indices array from sparse matrix.
+    slices_indptr : np.ndarray
+        Index pointer array from sparse matrix.
+    feature_offset_start : np.ndarray
+        Start offset for each feature.
+    feature_offset_end : np.ndarray
+        End offset for each feature.
+    feature_domains : np.ndarray
+        Domain size for each feature.
+
+    Returns
+    -------
+    np.ndarray
+        Array of unique IDs for each slice.
+    """
+    n_slices = len(slices_indptr) - 1
+    n_features = len(feature_offset_start)
+    slice_ids = np.zeros(n_slices, dtype=np.float64)
+
+    dom = feature_domains + 1
+
+    for i in range(n_slices):
+        start_idx = slices_indptr[i]
+        end_idx = slices_indptr[i + 1]
+
+        slice_id = 0.0
+        for j in range(start_idx, end_idx):
+            col = slices_indices[j]
+            val = slices_data[j]
+
+            if val == 0:
+                continue
+
+            for f in range(n_features):
+                if feature_offset_start[f] <= col < feature_offset_end[f]:
+                    offset = col - feature_offset_start[f]
+                    multiplier = 1.0
+                    for k in range(f + 1, n_features):
+                        multiplier *= dom[k]
+                    slice_id += (offset + 1) * multiplier
+                    break
+
+        slice_ids[i] = slice_id
+
+    return slice_ids

sliceline/slicefinder.py

@@ -2,20 +2,63 @@
 The slicefinder module implements the Slicefinder class.
 """
 
+from __future__ import annotations
+
 import logging
-from typing import Tuple, Union
+import warnings
+from typing import Any
 
 import numpy as np
+import numpy.typing as npt
 from scipy import sparse as sp
 from scipy.stats import rankdata
 from sklearn.base import BaseEstimator, TransformerMixin
 from sklearn.preprocessing import OneHotEncoder
-from sklearn.utils.validation import check_is_fitted
+from sklearn.utils.validation import _check_feature_names, check_is_fitted
 
 from sliceline.validation import check_array, check_X_e
 
+ArrayLike = npt.ArrayLike
+NDArray = npt.NDArray[Any]
+
 logger = logging.getLogger(__name__)
-logging.basicConfig(level=logging.INFO)
+
+# Numba availability detection
+try:
+    from sliceline._numba_ops import (
+        compute_slice_ids_numba,
+        score_slices_numba,
+        score_ub_batch_numba,
+    )
+
+    NUMBA_AVAILABLE = True
+except ImportError:
+    NUMBA_AVAILABLE = False
+    score_slices_numba = None
+    score_ub_batch_numba = None
+    compute_slice_ids_numba = None
+
+
+def is_numba_available() -> bool:
+    """Check if numba is available for acceleration.
+
+    Returns
+    -------
+    bool
+        True if numba is installed and can be used for acceleration.
+    """
+    return NUMBA_AVAILABLE
+
+
+def _warn_numba_not_available() -> None:
+    """Issue a warning if numba is not available."""
+    warnings.warn(
+        "Numba not available. Install with: pip install numba\n"
+        "Or: pip install sliceline[optimized]\n"
+        "Performance will be 5-50x slower without Numba optimization.",
+        UserWarning,
+        stacklevel=3,
+    )
 
 
 class Slicefinder(BaseEstimator, TransformerMixin):
@@ -93,9 +136,9 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         alpha: float = 0.6,
         k: int = 1,
         max_l: int = 4,
-        min_sup: Union[int, float] = 10,
+        min_sup: int | float = 10,
         verbose: bool = True,
-    ):
+    ) -> None:
         self.alpha = alpha
         self.k = k
         self.max_l = max_l
@@ -105,13 +148,25 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         self._one_hot_encoder = self._top_slices_enc = None
         self.top_slices_ = self.top_slices_statistics_ = None
         self.average_error_ = None
+        self._min_sup_actual = min_sup
 
         if self.verbose:
             logger.setLevel(logging.DEBUG)
         else:
             logger.setLevel(logging.INFO)
 
-    def _check_params(self):
+        # Warn user once if Numba optimization is not available
+        if not NUMBA_AVAILABLE and verbose:
+            warnings.warn(
+                "Numba JIT optimization not available. "
+                "Install with 'pip install sliceline[optimized]' "
+                "for 5-50x performance improvements on scoring operations. "
+                "See https://github.com/DataDome/sliceline for details.",
+                UserWarning,
+                stacklevel=2,
+            )
+
+    def _check_params(self) -> None:
         """Check transformer parameters."""
         if not 0 < self.alpha <= 1:
             raise ValueError(f"Invalid 'alpha' parameter: {self.alpha}")
@@ -127,7 +182,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         ):
             raise ValueError(f"Invalid 'min_sup' parameter: {self.min_sup}")
 
-    def _check_top_slices(self):
+    def _check_top_slices(self) -> None:
         """Check if slices have been found."""
         # Check if fit has been called
         check_is_fitted(self)
@@ -136,7 +191,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         if self.top_slices_.size == 0:
             raise ValueError("No transform: Sliceline did not find any slice.")
 
-    def fit(self, X, errors):
+    def fit(self, X: ArrayLike, errors: ArrayLike) -> Slicefinder:
         """Search for slice(s) on `X` based on `errors`.
 
         Parameters
@@ -155,20 +210,22 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         """
         self._check_params()
 
-        # Update min_sup for a fraction of the input dataset size
+        # Compute actual min_sup value (convert fraction to count if needed)
         if 0 < self.min_sup < 1:
-            self.min_sup = int(self.min_sup * len(X))
+            self._min_sup_actual = int(self.min_sup * len(X))
+        else:
+            self._min_sup_actual = self.min_sup
 
         # Check that X and e have correct shape
         X_array, errors = check_X_e(X, errors, y_numeric=True)
 
-        self._check_feature_names(X, reset=True)
+        _check_feature_names(self, X, reset=True)
 
         self._search_slices(X_array, errors)
 
         return self
 
-    def transform(self, X):
+    def transform(self, X: ArrayLike) -> NDArray:
         """Generate slices masks for `X`.
 
         Parameters
@@ -191,7 +248,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
         return slices_masks.T
 
-    def get_slice(self, X, slice_index: int):
+    def get_slice(self, X: ArrayLike, slice_index: int) -> NDArray:
         """Filter `X` samples according to the `slice_index`-th slice.
 
         Parameters
@@ -217,7 +274,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
         return X[np.where(slices_masks[slice_index])[0], :]
 
-    def get_feature_names_out(self):
+    def get_feature_names_out(self) -> NDArray:
         """Get output feature names for transformation.
 
         Returns
@@ -232,7 +289,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
         return np.array(feature_names, dtype=object)
 
-    def _get_slices_masks(self, X):
+    def _get_slices_masks(self, X: NDArray) -> NDArray:
         """Private utilities function generating slices masks for `X`."""
         X_encoded = self._one_hot_encoder.transform(X)
 
@@ -248,33 +305,51 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         return slices_masks
 
     @property
-    def _n_features_out(self):
+    def _n_features_out(self) -> int:
         """Number of transformed output features."""
         return self.top_slices_.shape[0]
 
     @staticmethod
-    def _dummify(array: np.ndarray, n_col_x_encoded: int) -> sp.csr_matrix:
+    def _dummify(array: NDArray, n_col_x_encoded: int) -> sp.csr_matrix:
         """Dummify `array` with respect to `n_col_x_encoded`.
-        Assumption: v does not contain any 0."""
-        assert (
-            0 not in array
-        ), "Modality 0 is not expected to be one-hot encoded."
-        one_hot_encoding = sp.lil_matrix(
-            (array.size, n_col_x_encoded), dtype=bool
+
+        Creates a sparse one-hot encoding matrix where each row corresponds
+        to an element in array and has a single True value in the column
+        specified by that element (adjusted for 1-based indexing).
+
+        Args:
+            array: 1-based indices to encode (must not contain 0)
+            n_col_x_encoded: Number of columns in output matrix
+
+        Returns:
+            Sparse CSR matrix of shape (len(array), n_col_x_encoded)
+
+        Raises:
+            ValueError: If array contains 0, which cannot be one-hot encoded.
+        """
+        if 0 in array:
+            raise ValueError(
+                "Modality 0 is not expected to be one-hot encoded."
+            )
+
+        # Direct CSR construction: 2-3x faster than lil_matrix approach
+        n = array.size
+        return sp.csr_matrix(
+            (np.ones(n, dtype=np.bool_), (np.arange(n), array - 1)),
+            shape=(n, n_col_x_encoded),
+            dtype=np.bool_,
         )
-        one_hot_encoding[np.arange(array.size), array - 1] = True
-        return one_hot_encoding.tocsr()
 
     def _maintain_top_k(
         self,
         slices: sp.csr_matrix,
-        statistics: np.ndarray,
+        statistics: NDArray,
         top_k_slices: sp.csr_matrix,
-        top_k_statistics: np.ndarray,
-    ) -> Tuple[sp.csr_matrix, np.ndarray]:
+        top_k_statistics: NDArray,
+    ) -> tuple[sp.csr_matrix, NDArray]:
         """Add new `slices` to `top_k_slices` and update the top-k slices."""
         # prune invalid min_sup and scores
-        valid_slices_mask = (statistics[:, 3] >= self.min_sup) & (
+        valid_slices_mask = (statistics[:, 3] >= self._min_sup_actual) & (
             statistics[:, 0] > 0
         )
         if np.sum(valid_slices_mask) != 0:
@@ -283,7 +358,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
                 statistics[valid_slices_mask],
             )
 
-            if (slices.shape[1] != top_k_slices.shape[1]) & (
+            if (slices.shape[1] != top_k_slices.shape[1]) and (
                 slices.shape[1] == 1
             ):
                 slices, statistics = slices.T, statistics.T
@@ -300,7 +375,14 @@ class Slicefinder(BaseEstimator, TransformerMixin):
                 slices[top_slices_bool],
                 statistics[top_slices_bool],
             )
-            top_slices_indices = np.argsort(-top_k_statistics[:, 0])
+            # Sort by score (descending), then lexicographically by slice representation
+            # to ensure deterministic ordering when scores are equal
+            scores = -top_k_statistics[:, 0]
+            slice_keys = tuple(
+                top_k_slices.toarray()[:, i]
+                for i in range(top_k_slices.shape[1])
+            )
+            top_slices_indices = np.lexsort(slice_keys[::-1] + (scores,))
             top_k_slices, top_k_statistics = (
                 top_k_slices[top_slices_indices],
                 top_k_statistics[top_slices_indices],
@@ -309,20 +391,35 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
     def _score_ub(
         self,
-        slice_sizes_ub: np.ndarray,
-        slice_errors_ub: np.ndarray,
-        max_slice_errors_ub: np.ndarray,
+        slice_sizes_ub: NDArray,
+        slice_errors_ub: NDArray,
+        max_slice_errors_ub: NDArray,
         n_col_x_encoded: int,
-    ) -> np.ndarray:
-        """Compute the upper-bound score for all the slices."""
+    ) -> NDArray:
+        """Compute the upper-bound score for all the slices.
+
+        Uses Numba JIT compilation when available for 5-10x speedup.
+        """
+        if NUMBA_AVAILABLE and score_ub_batch_numba is not None:
+            return score_ub_batch_numba(
+                slice_sizes_ub.astype(np.float64),
+                slice_errors_ub.astype(np.float64),
+                max_slice_errors_ub.astype(np.float64),
+                n_col_x_encoded,
+                float(self._min_sup_actual),
+                self.alpha,
+                self.average_error_,
+            )
+
+        # Fallback to NumPy implementation
         # Since slice_scores is either monotonically increasing or decreasing, we
         # probe interesting points of slice_scores in the interval [min_sup, ss],
         # and compute the maximum to serve as the upper bound
         potential_solutions = np.column_stack(
             (
-                self.min_sup * np.ones(slice_sizes_ub.shape[0]),
+                self._min_sup_actual * np.ones(slice_sizes_ub.shape[0]),
                 np.maximum(
-                    slice_errors_ub / max_slice_errors_ub, self.min_sup
+                    slice_errors_ub / max_slice_errors_ub, self._min_sup_actual
                 ),
                 slice_sizes_ub,
             )
@@ -346,7 +443,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         return slice_scores_ub
 
     @staticmethod
-    def _analyse_top_k(top_k_statistics: np.ndarray) -> tuple:
+    def _analyse_top_k(top_k_statistics: NDArray) -> tuple[float, float]:
         """Get the maximum and the minimum slices scores."""
         max_slice_scores = min_slice_scores = -np.inf
         if top_k_statistics.shape[0] > 0:
@@ -358,23 +455,40 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
     def _score(
         self,
-        slice_sizes: np.ndarray,
-        slice_errors: np.ndarray,
+        slice_sizes: NDArray,
+        slice_errors: NDArray,
         n_row_x_encoded: int,
-    ) -> np.ndarray:
-        """Compute the score for all the slices."""
-        slice_scores = self.alpha * (
-            (slice_errors / slice_sizes) / self.average_error_ - 1
-        ) - (1 - self.alpha) * (n_row_x_encoded / slice_sizes - 1)
+    ) -> NDArray:
+        """Compute the score for all the slices.
+
+        Uses Numba JIT compilation when available for 5-10x speedup.
+        """
+        if NUMBA_AVAILABLE and score_slices_numba is not None:
+            # Ensure inputs are float64 for numba
+            sizes = np.asarray(slice_sizes, dtype=np.float64)
+            errors = np.asarray(slice_errors, dtype=np.float64)
+            return score_slices_numba(
+                sizes,
+                errors,
+                n_row_x_encoded,
+                self.alpha,
+                self.average_error_,
+            )
+
+        # Fallback to NumPy implementation
+        with np.errstate(divide="ignore", invalid="ignore"):
+            slice_scores = self.alpha * (
+                (slice_errors / slice_sizes) / self.average_error_ - 1
+            ) - (1 - self.alpha) * (n_row_x_encoded / slice_sizes - 1)
         return np.nan_to_num(slice_scores, nan=-np.inf)
 
     def _eval_slice(
         self,
         x_encoded: sp.csr_matrix,
-        errors: np.ndarray,
+        errors: NDArray,
         slices: sp.csr_matrix,
         level: int,
-    ) -> np.ndarray:
+    ) -> NDArray:
         """Compute several statistics for all the slices."""
         slice_candidates = x_encoded @ slices.T == level
         slice_sizes = slice_candidates.sum(axis=0).A[0]
@@ -397,8 +511,8 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         self,
         x_encoded: sp.csr_matrix,
         n_col_x_encoded: int,
-        errors: np.ndarray,
-    ) -> Tuple[sp.csr_matrix, np.ndarray]:
+        errors: NDArray,
+    ) -> tuple[sp.csr_matrix, NDArray]:
         """Initialise 1-slices, i.e. slices with one predicate."""
         slice_sizes = x_encoded.sum(axis=0).A[0]
         slice_errors = errors @ x_encoded
@@ -409,7 +523,9 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         )
 
         # working set of active slices (#attr x #slices) and top-k
-        valid_slices_mask = (slice_sizes >= self.min_sup) & (slice_errors > 0)
+        valid_slices_mask = (slice_sizes >= self._min_sup_actual) & (
+            slice_errors > 0
+        )
         attr = np.arange(1, n_col_x_encoded + 1)[valid_slices_mask]
         slice_sizes = slice_sizes[valid_slices_mask]
         slice_errors = slice_errors[valid_slices_mask]
@@ -427,18 +543,18 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         n_col_dropped = n_col_x_encoded - sum(valid_slices_mask)
         logger.debug(
             "Dropping %i/%i features below min_sup = %i."
-            % (n_col_dropped, n_col_x_encoded, self.min_sup)
+            % (n_col_dropped, n_col_x_encoded, self._min_sup_actual)
         )
 
         return slices, statistics
 
     def _get_pruned_s_r(
-        self, slices: sp.csr_matrix, statistics: np.ndarray
-    ) -> Tuple[sp.csr_matrix, np.ndarray]:
+        self, slices: sp.csr_matrix, statistics: NDArray
+    ) -> tuple[sp.csr_matrix, NDArray]:
         """Prune invalid slices.
         Do not affect overall pruning effectiveness due to handling of missing parents.
         """
-        valid_slices_mask = (statistics[:, 3] >= self.min_sup) & (
+        valid_slices_mask = (statistics[:, 3] >= self._min_sup_actual) & (
             statistics[:, 1] > 0
         )
         return slices[valid_slices_mask], statistics[valid_slices_mask]
@@ -446,22 +562,53 @@ class Slicefinder(BaseEstimator, TransformerMixin):
     @staticmethod
     def _join_compatible_slices(
         slices: sp.csr_matrix, level: int
-    ) -> np.ndarray:
-        """Join compatible slices according to `level`."""
+    ) -> sp.csr_matrix:
+        """Join compatible slices keeping sparse format when beneficial.
+
+        Returns a sparse boolean matrix where entry (i,j) is True if slices
+        i and j are compatible for joining at the given level. Only upper
+        triangular entries (i < j) are populated.
+
+        For level==2 (looking for disjoint slices), uses dense format since
+        most pairs are compatible. For higher levels, keeps sparse format.
+        """
+        n_slices = slices.shape[0]
+        if n_slices == 0:
+            return sp.csr_matrix((0, 0), dtype=np.bool_)
+
         slices_int = slices.astype(int)
-        # Here we can't use the .A shorthand because it is not
-        # implemented in all scipy versions for coo_matrix objects
-        join = (slices_int @ slices_int.T).toarray() == level - 2
-        return np.triu(join, 1) * join
+        join_counts = slices_int @ slices_int.T
+
+        if level == 2:
+            # For level 2, we're looking for pairs with dot product == 0
+            # Most pairs will match, so dense is more efficient
+            join_dense = join_counts.toarray() == 0
+        else:
+            # For higher levels, most pairs won't match, so sparse is better
+            # Use dense conversion for smaller matrices to ensure consistent ordering
+            # This matches the original behavior and ensures deterministic results
+            join_dense = join_counts.toarray() == level - 2
+
+        join_upper = np.triu(join_dense, 1)
+        rows, cols = np.where(join_upper)
+        return sp.csr_matrix(
+            (np.ones(len(rows), dtype=np.bool_), (rows, cols)),
+            shape=join_counts.shape,
+            dtype=np.bool_,
+        )
 
     @staticmethod
     def _combine_slices(
         slices: sp.csr_matrix,
-        statistics: np.ndarray,
-        compatible_slices: np.ndarray,
-    ) -> Tuple[sp.csr_matrix, np.ndarray, np.ndarray, np.ndarray]:
-        """Combine slices by exploiting parents node statistics."""
-        parent_1_idx, parent_2_idx = np.where(compatible_slices == 1)
+        statistics: NDArray,
+        compatible_slices: sp.csr_matrix,
+    ) -> tuple[sp.csr_matrix, NDArray, NDArray, NDArray]:
+        """Combine slices by exploiting parents node statistics.
+
+        Works with sparse compatible_slices matrix returned by
+        _join_compatible_slices.
+        """
+        parent_1_idx, parent_2_idx = compatible_slices.nonzero()
         pair_candidates = slices[parent_1_idx] + slices[parent_2_idx]
 
         slice_errors = np.minimum(
@@ -477,13 +624,13 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
     @staticmethod
     def _prune_invalid_self_joins(
-        feature_offset_start: np.ndarray,
-        feature_offset_end: np.ndarray,
+        feature_offset_start: NDArray,
+        feature_offset_end: NDArray,
         pair_candidates: sp.csr_matrix,
-        slice_sizes: np.ndarray,
-        slice_errors: np.ndarray,
-        max_slice_errors: np.ndarray,
-    ) -> Tuple[sp.csr_matrix, np.ndarray, np.ndarray, np.ndarray]:
+        slice_sizes: NDArray,
+        slice_errors: NDArray,
+        max_slice_errors: NDArray,
+    ) -> tuple[sp.csr_matrix, NDArray, NDArray, NDArray]:
         """Prune invalid self joins (>1 bit per feature)."""
         valid_slices_mask = np.full(pair_candidates.shape[0], True)
         for start, end in zip(feature_offset_start, feature_offset_end):
@@ -500,12 +647,26 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
     @staticmethod
     def _prepare_deduplication_and_pruning(
-        feature_offset_start: np.ndarray,
-        feature_offset_end: np.ndarray,
-        feature_domains: np.ndarray,
+        feature_offset_start: NDArray,
+        feature_offset_end: NDArray,
+        feature_domains: NDArray,
         pair_candidates: sp.csr_matrix,
-    ) -> np.ndarray:
-        """Prepare IDs for deduplication and pruning."""
+    ) -> NDArray:
+        """Prepare IDs for deduplication and pruning.
+
+        Uses Numba JIT compilation when available for 10-50x speedup.
+        """
+        if NUMBA_AVAILABLE and compute_slice_ids_numba is not None:
+            return compute_slice_ids_numba(
+                pair_candidates.data.astype(np.float64),
+                pair_candidates.indices.astype(np.int64),
+                pair_candidates.indptr.astype(np.int64),
+                feature_offset_start.astype(np.int64),
+                feature_offset_end.astype(np.int64),
+                feature_domains.astype(np.float64),
+            )
+
+        # Fallback to Python implementation
         ids = np.zeros(pair_candidates.shape[0])
         dom = feature_domains + 1
         for j, (start, end) in enumerate(
@@ -525,18 +686,18 @@ class Slicefinder(BaseEstimator, TransformerMixin):
     def _get_pair_candidates(
         self,
         slices: sp.csr_matrix,
-        statistics: np.ndarray,
-        top_k_statistics: np.ndarray,
+        statistics: NDArray,
+        top_k_statistics: NDArray,
         level: int,
         n_col_x_encoded: int,
-        feature_domains: np.ndarray,
-        feature_offset_start: np.ndarray,
-        feature_offset_end: np.ndarray,
+        feature_domains: NDArray,
+        feature_offset_start: NDArray,
+        feature_offset_end: NDArray,
     ) -> sp.csr_matrix:
         """Compute and prune plausible slices candidates."""
         compatible_slices = self._join_compatible_slices(slices, level)
 
-        if np.sum(compatible_slices) == 0:
+        if compatible_slices.nnz == 0:
             return sp.csr_matrix(np.empty((0, slices.shape[1])))
 
         (
@@ -596,7 +757,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
         # Seems to be always fully True
         # Due to maintain_top_k that apply slice_sizes filter
-        pruning_sizes = slice_sizes >= self.min_sup
+        pruning_sizes = slice_sizes >= self._min_sup_actual
 
         _, min_slice_scores = self._analyse_top_k(top_k_statistics)
 
@@ -606,14 +767,14 @@ class Slicefinder(BaseEstimator, TransformerMixin):
 
     def _search_slices(
         self,
-        input_x: np.ndarray,
-        errors: np.ndarray,
+        input_x: NDArray,
+        errors: NDArray,
     ) -> None:
         """Main function of the SliceLine algorithm."""
         # prepare offset vectors and one-hot encoded input_x
         self._one_hot_encoder = OneHotEncoder(handle_unknown="ignore")
         x_encoded = self._one_hot_encoder.fit_transform(input_x)
-        feature_domains: np.ndarray = np.array(
+        feature_domains: NDArray = np.array(
             [len(sub_array) for sub_array in self._one_hot_encoder.categories_]
         )
         feature_offset_end = np.cumsum(feature_domains)
@@ -650,8 +811,8 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         min_condition = min(input_x.shape[1], self.max_l)
         while (
             (slices.shape[0] > 0)
-            & (slices.sum() > 0)
-            & (level < min_condition)
+            and (slices.sum() > 0)
+            and (level < min_condition)
         ):
             level += 1
 
@@ -687,7 +848,8 @@ class Slicefinder(BaseEstimator, TransformerMixin):
                 top_k_statistics
             )
             valid = np.sum(
-                (statistics[:, 3] >= self.min_sup) & (statistics[:, 1] > 0)
+                (statistics[:, 3] >= self._min_sup_actual)
+                & (statistics[:, 1] > 0)
             )
             logger.debug(
                 " -- valid slices after eval: %s/%i" % (valid, slices.shape[0])
@@ -723,7 +885,7 @@ class Slicefinder(BaseEstimator, TransformerMixin):
         ]
         self.top_slices_statistics_ = [
             {
-                stat_name: stat_value
+                stat_name: float(stat_value)
                 for stat_value, stat_name in zip(statistic, statistics_names)
             }
             for statistic in top_k_statistics

sliceline/validation.py

@@ -19,8 +19,11 @@ from contextlib import suppress
 import numpy as np
 import scipy.sparse as sp
 
-# mypy error: Module 'numpy.core.numeric' has no attribute 'ComplexWarning'
-from numpy.core.numeric import ComplexWarning  # type: ignore
+# ComplexWarning moved from numpy.core.numeric to numpy.exceptions in NumPy 2.0
+try:
+    from numpy.exceptions import ComplexWarning
+except ImportError:
+    from numpy.core.numeric import ComplexWarning  # type: ignore
 from sklearn._config import get_config as _get_config
 from sklearn.exceptions import DataConversionWarning
 from sklearn.utils.fixes import _object_dtype_isnan

{sliceline-0.2.18.dist-info → sliceline-0.3.0.dist-info}/METADATA

@@ -1,21 +1,41 @@
-Metadata-Version: 2.3
+Metadata-Version: 2.4
 Name: sliceline
-Version: 0.2.18
-Summary: ✂️ Fast slice finding for Machine Learning model debugging.
-License: BSD-3-Clause
+Version: 0.3.0
+Summary: Fast slice finding for Machine Learning model debugging.
+Project-URL: Homepage, https://github.com/DataDome/sliceline
+Project-URL: Documentation, https://sliceline.readthedocs.io/en/stable/
+Project-URL: Repository, https://github.com/DataDome/sliceline
 Author: Antoine de Daran
-Requires-Python: >=3.10,<4
+License-Expression: BSD-3-Clause
+License-File: LICENSE
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
 Classifier: License :: OSI Approved :: BSD License
+Classifier: Operating System :: OS Independent
 Classifier: Programming Language :: Python :: 3
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
-Classifier: Programming Language :: Python :: 3.13
-Requires-Dist: numpy (>=1.25,<2.0)
-Requires-Dist: scikit-learn (>=1.5.0,<2.0.0)
-Requires-Dist: scipy (>=1.12,<2.0)
-Project-URL: Documentation, https://sliceline.readthedocs.io/en/stable/
-Project-URL: Repository, https://github.com/DataDome/sliceline
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: <4,>=3.10
+Requires-Dist: numpy<3,>=1.25
+Requires-Dist: scikit-learn<2,>=1.6.0
+Requires-Dist: scipy<2,>=1.12
+Provides-Extra: dev
+Requires-Dist: jupyter>=1.0.0; extra == 'dev'
+Requires-Dist: matplotlib>=3.9; extra == 'dev'
+Requires-Dist: nbconvert>=7.0.0; extra == 'dev'
+Requires-Dist: optbinning>=0.15.0; extra == 'dev'
+Requires-Dist: pandas>=2.1.1; extra == 'dev'
+Requires-Dist: pytest-benchmark>=4.0.0; extra == 'dev'
+Requires-Dist: pytest-cov>=3.0.0; extra == 'dev'
+Requires-Dist: pytest>=7.2.0; extra == 'dev'
+Requires-Dist: ruff>=0.9.0; extra == 'dev'
+Requires-Dist: sphinx-rtd-theme>=3.0.0; extra == 'dev'
+Requires-Dist: sphinx>=8.0.0; extra == 'dev'
+Provides-Extra: optimized
+Requires-Dist: numba>=0.60.0; extra == 'optimized'
 Description-Content-Type: text/x-rst
 
 Sliceline
@@ -85,6 +105,47 @@ Or, through SSH:
 
    pip install git+ssh://git@github.com/datadome/sliceline.git --upgrade
 
+⚡ Performance Optimization
+---------------------------
+
+Sliceline includes optional Numba JIT compilation for **5-50x performance improvements** on scoring operations.
+
+**Quick Installation:**
+
+.. code:: sh
+
+   # With optimization support
+   pip install sliceline[optimized]
+
+**Benefits:**
+
+- 5-6x faster scoring operations
+- 1.4-4.5x faster overall fit() performance
+- Up to 17% memory reduction on large datasets
+- Automatic fallback to pure NumPy if Numba not available
+
+**System Requirements:**
+
+Numba requires LLVM to be installed:
+
+.. code:: sh
+
+   # macOS
+   brew install llvm
+
+   # Linux (Ubuntu/Debian)
+   sudo apt-get install llvm
+
+**Verify Optimization:**
+
+.. code:: python
+
+   from sliceline import is_numba_available
+
+   print("Numba available:", is_numba_available())
+
+See the `performance benchmarks <https://github.com/DataDome/sliceline/tree/main/benchmarks>`__ for detailed metrics.
+
 🔗 Useful links
 ---------------
 
@@ -116,4 +177,3 @@ if you want to bring modifications to the code base.
 ----------
 
 Sliceline is free and open-source software licensed under the `3-clause BSD license <https://github.com/DataDome/sliceline/blob/main/LICENSE>`__.
-

sliceline-0.3.0.dist-info/RECORD

@@ -0,0 +1,8 @@
+sliceline/__init__.py,sha256=6BE45x-4OhgXqkPRBmvXsTbfnZN5YzAUx8Wu8RIfWjE,106
+sliceline/_numba_ops.py,sha256=34IYOumWCfBMBpaNj5OxWcwJk2mFcnJTyKMmgTIjcIc,6455
+sliceline/slicefinder.py,sha256=umNscHR24iVU7C3kLTIkySMMAteD562vGIdcxHe3qLY,32042
+sliceline/validation.py,sha256=pydiTHlS6f1iBtlIqATLVHimhoyZDKTMrDjQIH2R9ks,30875
+sliceline-0.3.0.dist-info/METADATA,sha256=_rAxmaiYiWTNgnZzFO-ijMJAl2vPd4z7B0XZLCn4oQU,5518
+sliceline-0.3.0.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+sliceline-0.3.0.dist-info/licenses/LICENSE,sha256=AbeN2ySrCt8VUJukqcQIYutROwZh3W2u0UU1d7EnbZs,1531
+sliceline-0.3.0.dist-info/RECORD,,

{sliceline-0.2.18.dist-info → sliceline-0.3.0.dist-info}/WHEEL

@@ -1,4 +1,4 @@
 Wheel-Version: 1.0
-Generator: poetry-core 2.1.1
+Generator: hatchling 1.28.0
 Root-Is-Purelib: true
 Tag: py3-none-any

sliceline-0.2.18.dist-info/RECORD

@@ -1,7 +0,0 @@
-sliceline/__init__.py,sha256=jEIUmQtv4W_eZuH63KQ8tAFoRZxyN3O8bRZ__FlMJr0,65
-sliceline/slicefinder.py,sha256=xuGsxGXtihkKNEokRmhycJ6aY-FPkkNjCsPQKTcPABg,26355
-sliceline/validation.py,sha256=-RkCpRdANNeaJyrdj7zFn4xs1X1xIXitKwRoL_B5EAk,30794
-sliceline-0.2.18.dist-info/LICENSE,sha256=AbeN2ySrCt8VUJukqcQIYutROwZh3W2u0UU1d7EnbZs,1531
-sliceline-0.2.18.dist-info/METADATA,sha256=JiKzXGFQfX7pOMhuhafNrtKxB57yTK7z91CySd3MD7I,3717
-sliceline-0.2.18.dist-info/WHEEL,sha256=XbeZDeTWKc1w7CSIyre5aMDU_-PohRwTQceYnisIYYY,88
-sliceline-0.2.18.dist-info/RECORD,,