deskit 1.2.2__tar.gz → 1.2.3__tar.gz

{deskit-1.2.2/src/deskit.egg-info → deskit-1.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deskit
-Version: 1.2.2
+Version: 1.2.3
 Summary: A Python library for Dynamic Ensemble Selection
 Author: Tikhon Vodyanov
 License-Expression: MIT

{deskit-1.2.2 → deskit-1.2.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "deskit"
-version = "1.2.2"
+version = "1.2.3"
 description = "A Python library for Dynamic Ensemble Selection"
 readme = "README.md"
 license = "MIT"

deskit-1.2.3/src/deskit/base/knnbase.py

@@ -0,0 +1,135 @@
+from deskit.base.base import BaseRouter
+from deskit.base.predictbase import PredictBase
+import numpy as np
+
+
+class KNNBase(PredictBase, BaseRouter):
+    """
+    Base for KNN-based DES algorithms.
+
+    Inheriting PredictBase gives every subclass the public
+    predict() and predict_weights() API automatically.
+    Subclasses must implement _weights_batch().
+    """
+
+    def __init__(self, metric, mode='max', neighbor_finder=None, task='classification'):
+        """
+        Parameters
+        ----------
+        metric : callable
+            Per-sample scoring function: (y_true, y_pred) -> float.
+        mode : str
+            'max' if higher scores are better, 'min' if lower.
+        neighbor_finder : NeighborFinder
+            Backend used for neighborhood queries.
+        """
+        self.metric          = metric
+        self.mode            = mode
+        self.model           = neighbor_finder
+        self.matrix          = None   # (n_val, n_models); higher is always better
+        self.models          = None   # ordered list of model names
+        self.task = task
+
+    def _compute_scores(self, y, preds):
+        """
+        Return a 1D array of per-sample metric scores.
+
+        preds may be 1D (scalar predictions) or 2D (probability arrays, one
+        row per sample).
+        """
+        preds = np.asarray(preds)
+        if preds.ndim == 2:
+            return np.array([self.metric(y[i], preds[i]) for i in range(len(y))])
+        return np.vectorize(self.metric)(y, preds)
+
+    def fit(self, features, y, preds_dict):
+        """
+        Build the score matrix and fit the neighbor index.
+
+        This method expects pre-validated numpy arrays.
+        """
+        self.models = list(preds_dict.keys())
+        n_val       = len(y)
+        n_models    = len(self.models)
+        self.matrix = np.zeros((n_val, n_models))
+
+        for j, name in enumerate(self.models):
+            scores = self._compute_scores(y, preds_dict[name])
+            self.matrix[:, j] = scores if self.mode == 'max' else -scores
+
+        self.model.fit(features)
+
+    def _kneighbors(self, x, k=None, loo=False):
+        """
+        Query the fitted neighbor index, with optional leave-one-out (LOO)
+        exclusion of each query point's own occurrence in the DSEL.
+
+        Set loo=True when ``x`` is (part of) the same data this model was
+        fit on -- e.g. while tuning k / threshold / temperature directly on
+        the DSEL -- so a point doesn't end up neighboring itself at distance
+        0, which would otherwise dominate the routing.
+
+        Parameters
+        ----------
+        x : np.ndarray, shape (batch, n_features)
+        k : int, optional
+            Neighborhood size. None defers to the finder's own default.
+        loo : bool
+            If True, query one extra neighbor per row and drop the
+            zero-distance match (the point itself) when present, so the
+            returned neighborhood still has the size a normal call would
+            have produced. Rows with no zero-distance match (e.g. ``x``
+            isn't actually part of the fitted DSEL) fall back to dropping
+            the farthest neighbor instead, so shapes stay consistent.
+
+        Returns
+        -------
+        distances, indices : np.ndarray, each shape (batch, k_eff)
+        """
+        if not loo:
+            return self.model.kneighbors(x, k=k)
+
+        # Different backends store their default k under different
+        # attribute names (n_neighbors vs k), so rather than guessing,
+        # resolve the effective k with one cheap probe call when it isn't
+        # given explicitly. Costs one extra query only in that case.
+        if k is None:
+            probe_distances, _ = self.model.kneighbors(x, k=k)
+            k = probe_distances.shape[1]
+
+        distances, indices = self.model.kneighbors(x, k=k + 1)
+        return _drop_self_match(distances, indices, k)
+
+
+def _drop_self_match(distances, indices, k, eps=1e-6):
+    """
+    Drop one zero-distance neighbor per row from (batch, k+1) neighbor
+    results, returning (batch, k) arrays.
+
+    A query point's own occurrence in the DSEL, if present, is always the
+    closest possible neighbor (distance 0 is the global minimum for any
+    proper distance metric), so it's identified per row as whichever
+    column is nearest, gated on that distance being ~0. Rows without a
+    zero-distance match (point not actually in the DSEL) drop the
+    farthest neighbor instead, so every row keeps exactly k entries.
+
+    Note: if the DSEL contains true duplicate feature rows, only one
+    occurrence is dropped per query point -- the duplicates remain valid,
+    distinct neighbors. Distance order within each row need not be
+    sorted; this works either way.
+    """
+    batch = distances.shape[0]
+    rows = np.arange(batch)
+
+    nearest_col = np.argmin(distances, axis=1)
+    is_self_match = distances[rows, nearest_col] < eps
+
+    farthest_col = np.argmax(distances, axis=1)
+    drop_col = np.where(is_self_match, nearest_col, farthest_col)
+
+    keep = np.ones(distances.shape, dtype=bool)
+    keep[rows, drop_col] = False
+
+    new_distances = distances[keep].reshape(batch, k)
+    new_indices = indices[keep].reshape(batch, k)
+    return new_distances, new_indices

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/base/predictbase.py

@@ -42,6 +42,11 @@ class PredictBase:
         **kwargs
             Additional per-call overrides forwarded to _weights_batch.
             Supported by most algorithms: ``threshold=<float>``.
+            KNN-based algorithms also support ``loo=<bool>`` (default
+            False): set True when X_test is (part of) the same data the
+            model was fit on -- e.g. while tuning hyperparameters on the
+            DSEL -- so each point's own occurrence is excluded from its
+            neighborhood instead of trivially matching itself at distance 0.
 
         Returns
         -------
@@ -79,7 +84,9 @@ class PredictBase:
         temperature : float, optional
             Forwarded to _weights_batch.
         **kwargs
-            Forwarded to _weights_batch (e.g. ``threshold=<float>``).
+            Forwarded to _weights_batch (e.g. ``threshold=<float>``, or
+            ``loo=<bool>`` for leave-one-out neighbor exclusion when
+            X_test is drawn from the fit DSEL -- see predict_weights).
 
         Returns
         -------

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/dewsi.py

@@ -38,6 +38,13 @@ class DEWSI(KNNBase):
     distance_metric : str
         Distance function to use for neighbor search. Default: 'euclidean'. See
         neighbors.list_distance_metrics() for all options and per-backend availability.
+
+    Notes
+    -----
+    predict() and predict_weights() accept an optional ``loo=True`` keyword
+    for hyperparameter tuning directly on the DSEL this model was fit on:
+    it excludes each query point's own occurrence from its neighborhood so
+    it doesn't trivially neighbor itself at distance 0.
     """
 
     def __init__(self, task, metric='mae', mode='min', k=10,
@@ -69,17 +76,23 @@ class DEWSI(KNNBase):
         )
         super().fit(features, y, preds_dict)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
+
+        loo : bool
+            Leave-one-out. Set True when x is (part of) the DSEL this model
+            was fit on -- e.g. while tuning hyperparameters -- so each
+            point's own occurrence is excluded from its neighborhood
+            instead of trivially matching itself at distance 0.
         """
         t  = temperature if temperature is not None else (
              self._temperature if self._temperature is not None else
              (0.5 if self.mode == 'min' else 1.0))
         th = threshold if threshold is not None else self.threshold
 
-        distances, indices = self.model.kneighbors(x, k=k)               # both (batch, k)
+        distances, indices = self._kneighbors(x, k=k, loo=loo)           # both (batch, k)
 
         # Inverse-distance-weighted average of each model's scores over K neighbors
         inv_dist   = 1.0 / np.maximum(distances, 1e-8)              # (batch, k)

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/dewsiv.py

@@ -84,7 +84,7 @@ class DEWSIV(KNNBase):
                 preds = np.asarray(preds_dict[name])
                 self._var_matrix[:, j] = np.vectorize(_signed_residual)(y, preds)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -94,7 +94,7 @@ class DEWSIV(KNNBase):
              (0.5 if self.mode == 'min' else 1.0))
         th = threshold if threshold is not None else self.threshold
 
-        distances, indices = self.model.kneighbors(x, k=k)                # both (batch, k)
+        distances, indices = self.model.kneighbors(x, k=k, loo=loo)                # both (batch, k)
 
         # Inverse-distance weights
         inv_dist   = 1.0 / np.maximum(distances, 1e-8)               # (batch, k)

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/dewst.py

@@ -86,7 +86,7 @@ class DEWST(KNNBase):
         )
         super().fit(features, y, preds_dict)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None, r2_threshold=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, r2_threshold=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -97,7 +97,7 @@ class DEWST(KNNBase):
         th    = threshold    if threshold    is not None else self.threshold
         r2_th = r2_threshold if r2_threshold is not None else self.r2_threshold
 
-        distances, indices = self.model.kneighbors(x, k=k)          # (batch, k)
+        distances, indices = self.model.kneighbors(x, k=k, loo=loo)          # (batch, k)
         k = distances.shape[1]
 
         # Inverse-distance weights

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/dewsu.py

@@ -65,7 +65,7 @@ class DEWSU(KNNBase):
         )
         super().fit(features, y, preds_dict)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -75,7 +75,7 @@ class DEWSU(KNNBase):
              (0.5 if self.mode == 'min' else 1.0))
         th = threshold if threshold is not None else self.threshold
 
-        _, indices = self.model.kneighbors(x, k=k)                        # (batch, k)
+        _, indices = self.model.kneighbors(x, k=k, loo=loo)                        # (batch, k)
 
         # Average each model's scores over the K neighbors
         avg_scores = self.matrix[indices].mean(axis=1)               # (batch, n_models)

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/dewsv.py

@@ -84,7 +84,7 @@ class DEWSV(KNNBase):
                 preds = np.asarray(preds_dict[name])
                 self._var_matrix[:, j] = np.vectorize(_signed_residual)(y, preds)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -94,7 +94,7 @@ class DEWSV(KNNBase):
              (0.5 if self.mode == 'min' else 1.0))
         th = threshold if threshold is not None else self.threshold
 
-        _, indices = self.model.kneighbors(x, k=k)                        # (batch, k)
+        _, indices = self.model.kneighbors(x, k=k, loo=loo)                        # (batch, k)
 
         # Uniform average of each model's scores over K neighbors
         neighbor_scores = self.matrix[indices]                        # (batch, k, n_models)

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/knorae.py

@@ -55,7 +55,7 @@ class KNORAE(KNNBase):
         )
         super().fit(features, y, preds_dict)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -64,7 +64,7 @@ class KNORAE(KNNBase):
         th       = threshold if threshold is not None else self.threshold
         n_models = len(self.models)
 
-        _, indices      = self.model.kneighbors(x, k=k)
+        _, indices      = self.model.kneighbors(x, k=k, loo=loo)
         k               = indices.shape[1]
         neighbor_scores = self.matrix[indices]                        # (batch, k, n_models)
 

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/knoraiu.py

@@ -56,7 +56,7 @@ class KNORAIU(KNNBase):
         )
         super().fit(features, y, preds_dict)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -64,7 +64,7 @@ class KNORAIU(KNNBase):
         """
         th = threshold if threshold is not None else self.threshold
 
-        distances, indices = self.model.kneighbors(x, k=k)                # both (batch, k)
+        distances, indices = self.model.kneighbors(x, k=k, loo=loo)                # both (batch, k)
         neighbor_scores    = self.matrix[indices]                     # (batch, k, n_models)
 
         # Normalize per neighbor: best model = 1.0, worst = 0.0

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/knorau.py

@@ -56,7 +56,7 @@ class KNORAU(KNNBase):
         )
         super().fit(features, y, preds_dict)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -64,7 +64,7 @@ class KNORAU(KNNBase):
         """
         th = threshold if threshold is not None else self.threshold
 
-        _, indices      = self.model.kneighbors(x, k=k)
+        _, indices      = self.model.kneighbors(x, k=k, loo=loo)
         neighbor_scores = self.matrix[indices]                        # (batch, k, n_models)
 
         # Normalize per neighbor: best model = 1.0, worst = 0.0

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/lwsei.py

@@ -74,7 +74,7 @@ class LWSEI(PredictBase):
         self._y_val = y
         self._finder.fit(features)
 
-    def _weights_batch(self, x, temperature=None, k=None, **kwargs):
+    def _weights_batch(self, x, temperature=None, k=None, loo=False, **kwargs):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -87,7 +87,7 @@ class LWSEI(PredictBase):
         n_models   = len(self.models)
         uniform    = np.full(n_models, 1.0 / n_models)
 
-        distances, indices = self._finder.kneighbors(x, k=k)              # (batch, k)
+        distances, indices = self._finder.kneighbors(x, k=k, loo=loo)              # (batch, k)
         weights_out        = np.empty((batch_size, n_models))
 
         for b in range(batch_size):

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/lwseu.py

@@ -74,7 +74,7 @@ class LWSEU(PredictBase):
         self._y_val = y
         self._finder.fit(features)
 
-    def _weights_batch(self, x, temperature=None, k=None, **kwargs):
+    def _weights_batch(self, x, temperature=None, k=None, loo=False, **kwargs):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -87,7 +87,7 @@ class LWSEU(PredictBase):
         n_models   = len(self.models)
         uniform    = np.full(n_models, 1.0 / n_models)
 
-        distances, indices = self._finder.kneighbors(x, k=k)              # (batch, k)
+        distances, indices = self._finder.kneighbors(x, k=k, loo=loo)              # (batch, k)
         weights_out        = np.empty((batch_size, n_models))
 
         for b in range(batch_size):

{deskit-1.2.2 → deskit-1.2.3}/src/deskit/des/ola.py

@@ -54,7 +54,7 @@ class OLA(KNNBase):
         if mat_max > mat_min:
             self.matrix = (self.matrix - mat_min) / (mat_max - mat_min)
 
-    def _weights_batch(self, x, temperature=None, threshold=None, k=None):
+    def _weights_batch(self, x, temperature=None, threshold=None, k=None, loo=False):
         """
         Core weight computation. x is a 2-D float64 numpy array (batch, n_features).
         Returns (batch, n_models) weight array.
@@ -63,7 +63,7 @@ class OLA(KNNBase):
         """
         batch_size = x.shape[0]
 
-        _, indices  = self.model.kneighbors(x, k=k)
+        _, indices  = self.model.kneighbors(x, k=k, loo=loo)
         avg_scores  = self.matrix[indices].mean(axis=1)               # (batch, n_models)
         best_indices = np.argmax(avg_scores, axis=1)
 

{deskit-1.2.2 → deskit-1.2.3/src/deskit.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deskit
-Version: 1.2.2
+Version: 1.2.3
 Summary: A Python library for Dynamic Ensemble Selection
 Author: Tikhon Vodyanov
 License-Expression: MIT

deskit-1.2.2/src/deskit/base/knnbase.py

@@ -1,60 +0,0 @@
-from deskit.base.base import BaseRouter
-from deskit.base.predictbase import PredictBase
-import numpy as np
-
-
-class KNNBase(PredictBase, BaseRouter):
-    """
-    Base for KNN-based DES algorithms.
-
-    Inheriting PredictBase gives every subclass the public
-    predict() and predict_weights() API automatically.
-    Subclasses must implement _weights_batch().
-    """
-
-    def __init__(self, metric, mode='max', neighbor_finder=None, task='classification'):
-        """
-        Parameters
-        ----------
-        metric : callable
-            Per-sample scoring function: (y_true, y_pred) -> float.
-        mode : str
-            'max' if higher scores are better, 'min' if lower.
-        neighbor_finder : NeighborFinder
-            Backend used for neighborhood queries.
-        """
-        self.metric          = metric
-        self.mode            = mode
-        self.model           = neighbor_finder
-        self.matrix          = None   # (n_val, n_models); higher is always better
-        self.models          = None   # ordered list of model names
-        self.task = task
-
-    def _compute_scores(self, y, preds):
-        """
-        Return a 1D array of per-sample metric scores.
-
-        preds may be 1D (scalar predictions) or 2D (probability arrays, one
-        row per sample).
-        """
-        preds = np.asarray(preds)
-        if preds.ndim == 2:
-            return np.array([self.metric(y[i], preds[i]) for i in range(len(y))])
-        return np.vectorize(self.metric)(y, preds)
-
-    def fit(self, features, y, preds_dict):
-        """
-        Build the score matrix and fit the neighbor index.
-
-        This method expects pre-validated numpy arrays.
-        """
-        self.models = list(preds_dict.keys())
-        n_val       = len(y)
-        n_models    = len(self.models)
-        self.matrix = np.zeros((n_val, n_models))
-
-        for j, name in enumerate(self.models):
-            scores = self._compute_scores(y, preds_dict[name])
-            self.matrix[:, j] = scores if self.mode == 'max' else -scores
-
-        self.model.fit(features)