Merge pull request #53 from gmrukwa/kdtree-percentile-init

Kdtree percentile init

.github/workflows/deploy.yml

@@ -12,7 +12,7 @@ on:
 env:
   MAJOR: ${{ 2 }}
   MINOR: ${{ 4 }}
-  FIXUP: ${{ 6 }}
+  FIXUP: ${{ 7 }}
   PACKAGE_INIT_FILE: ${{ 'divik/__init__.py' }}
   PACKAGE_INIT_FILE_VERSION_LINE: ${{ 1 }}
   PACKAGE_SETUP_FILE: ${{ 'setup.py' }}

README.md

@@ -40,7 +40,7 @@ docker pull gmrukwa/divik
 To install specific version, you can specify it in the command, e.g.:
 
 ```bash
-docker pull gmrukwa/divik:2.4.6
+docker pull gmrukwa/divik:2.4.7
 ```
 
 ## Python package
@@ -79,7 +79,7 @@ pip install divik
 or any stable tagged version, e.g.:
 
 ```bash
-pip install divik==2.4.6
+pip install divik==2.4.7
 ```
 
 If you want to have compatibility with

divik/__init__.py

@@ -1,4 +1,4 @@
-__version__ = '2.4.6'
+__version__ = '2.4.7'
 
 from ._summary import plot, reject_split
 

divik/cluster/_kmeans/_core.py

@@ -12,6 +12,7 @@
     ExtremeInitialization,
     PercentileInitialization,
     KDTreeInitialization,
+    KDTreePercentileInitialization,
 )
 from divik.core import (
     normalize_rows,
@@ -157,6 +158,8 @@ def _parse_initialization(name: str, distance: str,
         return ExtremeInitialization(distance)
     if name == 'kdtree':
         return KDTreeInitialization(distance, leaf_size)
+    if name == 'kdtree_percentile':
+        return KDTreePercentileInitialization(distance, leaf_size, percentile)
     raise ValueError('Unknown initialization: {0}'.format(name))
 
 
@@ -174,7 +177,7 @@ class KMeans(BaseEstimator, ClusterMixin, TransformerMixin):
     distance : str, optional, default: 'euclidean'
         Distance measure. One of the distances supported by scipy package.
 
-    init : {'percentile', 'extreme' or 'kdtree'}
+    init : {'percentile', 'extreme', 'kdtree', 'kdtree_percentile'}
         Method for initialization, defaults to 'percentile':
 
         'percentile' : selects initial cluster centers for k-mean
@@ -188,6 +191,10 @@ class KMeans(BaseEstimator, ClusterMixin, TransformerMixin):
         'kdtree': selects initial cluster centers for k-mean clustering
         starting from centroids of KD-Tree boxes
 
+        'kdtree_percentile': selects initial cluster centers for k-means
+        clustering starting from centroids of KD-Tree boxes containing
+        specified percentile. This should be more robust against outliers.
+
     percentile : float, default: 95.0
         Specifies the starting percentile for 'percentile' initialization.
         Must be within range [0.0, 100.0]. At 100.0 it is equivalent to

divik/cluster/_kmeans/_initialization.py

@@ -214,3 +214,49 @@ def __call__(self, data: Data, number_of_centroids: int) -> Centroids:
             centroids[i] = box_centroids[np.argmax(distances)]
 
         return centroids
+
+
+class KDTreePercentileInitialization(Initialization):
+    """Initializes k-means by picking extreme KDTree box"""
+    def __init__(self, distance: str, leaf_size: Union[int, float] = 0.01,
+                 percentile: float=99.):
+        assert 0 <= percentile <= 100, percentile
+        self.distance = distance
+        self.leaf_size = leaf_size
+        self.percentile = percentile
+
+    def _get_percentile_idx(self, distances, weights) -> int:
+        idx = np.argsort(distances)
+        over_percentile = np.cumsum(weights[idx]) >= self.percentile
+        first_over = np.flatnonzero(over_percentile)[0]
+        return idx[first_over]
+
+    def __call__(self, data: Data, number_of_centroids: int) -> Centroids:
+        """Generate initial centroids for k-means algorithm"""
+        _validate(data, number_of_centroids)
+        leaf_size = self.leaf_size
+        if isinstance(leaf_size, float):
+            if 0 <= leaf_size <= 1:
+                leaf_size = max(int(leaf_size * data.shape[0]), 1)
+            else:
+                raise ValueError('leaf_size must be between 0 and 1 when float')
+        tree = make_tree(data, leaf_size=leaf_size)
+        leaves = get_leaves(tree)
+        box_centroids = np.vstack([l.centroid for l in leaves])
+        box_weights = np.array([l.count for l in leaves])
+        normalized_weights = box_weights / np.sum(box_weights)
+
+        residuals = _find_residuals(box_centroids, box_weights)
+        centroids = np.nan * np.zeros((number_of_centroids, data.shape[1]))
+        idx = self._get_percentile_idx(residuals, normalized_weights)
+        centroids[0] = box_centroids[idx]
+
+        distances = np.inf * np.ones((box_centroids.shape[0], ))
+        for i in range(1, number_of_centroids):
+            current_distance = dist.cdist(
+                box_centroids, centroids[np.newaxis, i - 1], self.distance)
+            distances[:] = np.minimum(current_distance.ravel(), distances)
+            idx = self._get_percentile_idx(distances, normalized_weights)
+            centroids[i] = box_centroids[idx]
+
+        return centroids

docs/instructions/installation.rst

@@ -14,7 +14,7 @@ To install latest stable version use::
 
 To install specific version, you can specify it in the command, e.g.::
 
-    docker pull gmrukwa/divik:2.4.6
+    docker pull gmrukwa/divik:2.4.7
 
 Python package
 --------------
@@ -31,7 +31,7 @@ package::
 
 or any stable tagged version, e.g.::
 
-    pip install divik==2.4.6
+    pip install divik==2.4.7
 
 If you want to have compatibility with
 `gin-config <https://github.com/google/gin-config>`_, you can install

setup.py

@@ -6,7 +6,7 @@
 import sys
 import numpy
 
-__version__ = '2.4.6'
+__version__ = '2.4.7'
 
 LINUX_OPTS = {
     'extra_link_args': [