SilhouetteVisualizer add support for more estimators (DistrictDataLabs#1294)

Signed-off-by: Benjamin Bengfort <[email protected]>
Co-authored-by: Benjamin Bengfort <[email protected]>
Co-authored-by: Larry Gray <[email protected]>

Author: 3 people

tests/baseline_images/test_cluster/test_silhouette/test_colormap_as_colors_silhouette.png

@@ -20,14 +20,15 @@
 import sys
 import pytest
 import matplotlib.pyplot as plt
+import numpy as np
 
 from sklearn.datasets import make_blobs
 from sklearn.cluster import KMeans, MiniBatchKMeans
+from sklearn.cluster import SpectralClustering, AgglomerativeClustering
 
 from unittest import mock
 from tests.base import VisualTestCase
 
-from yellowbrick.datasets import load_nfl
 from yellowbrick.cluster.silhouette import SilhouetteVisualizer, silhouette_visualizer
 
 
@@ -53,7 +54,6 @@ def test_integrated_kmeans_silhouette(self):
             n_samples=1000, n_features=12, centers=8, shuffle=False, random_state=0
         )
 
-       
         fig = plt.figure()
         ax = fig.add_subplot()
 
@@ -62,7 +62,6 @@ def test_integrated_kmeans_silhouette(self):
         visualizer.finalize()
 
         self.assert_images_similar(visualizer, remove_legend=True)
-        
 
     @pytest.mark.xfail(sys.platform == "win32", reason="images not close on windows")
     def test_integrated_mini_batch_kmeans_silhouette(self):
@@ -84,7 +83,6 @@ def test_integrated_mini_batch_kmeans_silhouette(self):
         visualizer.finalize()
 
         self.assert_images_similar(visualizer, remove_legend=True)
-        
 
     @pytest.mark.skip(reason="no negative silhouette example available yet")
     def test_negative_silhouette_score(self):
@@ -103,7 +101,6 @@ def test_colormap_silhouette(self):
             n_samples=1000, n_features=12, centers=8, shuffle=False, random_state=0
         )
 
-        
         fig = plt.figure()
         ax = fig.add_subplot()
 
@@ -138,7 +135,7 @@ def test_colors_silhouette(self):
         visualizer.finalize()
 
         self.assert_images_similar(visualizer, remove_legend=True)
-    
+
     def test_colormap_as_colors_silhouette(self):
         """
         Test no exceptions for modifying the colors in a silhouette visualizer
@@ -162,7 +159,7 @@ def test_colormap_as_colors_silhouette(self):
             3.2 if sys.platform == "win32" else 0.01
         )  # Fails on AppVeyor with RMS 3.143
         self.assert_images_similar(visualizer, remove_legend=True, tol=tol)
-    
+
     def test_quick_method(self):
         """
         Test the quick method producing a valid visualization
@@ -177,29 +174,44 @@ def test_quick_method(self):
 
         self.assert_images_similar(oz)
 
-    @pytest.mark.xfail(
-        reason="""third test fails with AssertionError: Expected fit
-        to be called once. Called 0 times."""
-    )
     def test_with_fitted(self):
         """
         Test that visualizer properly handles an already-fitted model
         """
-        X, y = load_nfl(return_dataset=True).to_numpy()
-
-        model = MiniBatchKMeans().fit(X, y)
+        X, y = make_blobs(
+            n_samples=100, n_features=5, centers=3, shuffle=False, random_state=112
+        )
+        model = MiniBatchKMeans().fit(X)
+        labels = model.predict(X)
 
         with mock.patch.object(model, "fit") as mockfit:
             oz = SilhouetteVisualizer(model)
-            oz.fit(X, y)
+            oz.fit(X)
             mockfit.assert_not_called()
 
         with mock.patch.object(model, "fit") as mockfit:
             oz = SilhouetteVisualizer(model, is_fitted=True)
-            oz.fit(X, y)
+            oz.fit(X)
             mockfit.assert_not_called()
 
-        with mock.patch.object(model, "fit") as mockfit:
+        with mock.patch.object(model, "fit_predict", return_value=labels) as mockfit:
             oz = SilhouetteVisualizer(model, is_fitted=False)
-            oz.fit(X, y)
-            mockfit.assert_called_once_with(X, y)
+            oz.fit(X)
+            mockfit.assert_called_once_with(X, None)
+
+    @pytest.mark.parametrize(
+        "model",
+        [SpectralClustering, AgglomerativeClustering],
+    )
+    def test_clusterer_without_predict(self, model):
+        """
+        Assert that clustering estimators that don't implement
+        a predict() method utilize fit_predict()
+        """
+        X = np.array([[1, 2], [1, 4], [1, 0], [4, 2], [4, 4], [4, 0]])
+        try:
+            visualizer = SilhouetteVisualizer(model(n_clusters=2))
+            visualizer.fit(X)
+            visualizer.finalize()
+        except AttributeError:
+            self.fail("could not use fit or fit_predict methods")

tests/baseline_images/test_cluster/test_silhouette/test_colormap_silhouette.png

@@ -23,6 +23,35 @@
 
 from sklearn.metrics import silhouette_score, silhouette_samples
 
+try:
+    from sklearn.metrics.pairwise import _VALID_METRICS
+except ImportError:
+    _VALID_METRICS = [
+        "cityblock",
+        "cosine",
+        "euclidean",
+        "l1",
+        "l2",
+        "manhattan",
+        "braycurtis",
+        "canberra",
+        "chebyshev",
+        "correlation",
+        "dice",
+        "hamming",
+        "jaccard",
+        "kulsinski",
+        "mahalanobis",
+        "minkowski",
+        "rogerstanimoto",
+        "russellrao",
+        "seuclidean",
+        "sokalmichener",
+        "sokalsneath",
+        "sqeuclidean",
+        "yule",
+    ]
+
 from yellowbrick.utils import check_fitted
 from yellowbrick.style import resolve_colors
 from yellowbrick.cluster.base import ClusteringScoreVisualizer
@@ -113,7 +142,6 @@ class SilhouetteVisualizer(ClusteringScoreVisualizer):
     """
 
     def __init__(self, estimator, ax=None, colors=None, is_fitted="auto", **kwargs):
-
         # Initialize the visualizer bases
         super(SilhouetteVisualizer, self).__init__(
             estimator, ax=ax, is_fitted=is_fitted, **kwargs
@@ -130,23 +158,47 @@ def __init__(self, estimator, ax=None, colors=None, is_fitted="auto", **kwargs):
     def fit(self, X, y=None, **kwargs):
         """
         Fits the model and generates the silhouette visualization.
+
+        Unlike other visualizers that use the score() method to draw the results, this
+        visualizer errs on visualizing on fit since this is when the clusters are
+        computed. This means that a predict call is required in fit (or a fit_predict)
+        in order to produce the visualization.
         """
-        # TODO: decide to use this method or the score method to draw.
-        # NOTE: Probably this would be better in score, but the standard score
-        # is a little different and I'm not sure how it's used.
 
+        # If the estimator is not fitted, fit it; then call predict to get the labels
+        # for computing the silhoutte score on. If the estimator is already fitted, then
+        # attempt to predict the labels, but if the estimator is stateless, fit and
+        # predict on the data specified. At the end of this block, no matter the fitted
+        # state of the estimator and the method, we should have cluster labels for X.
         if not check_fitted(self.estimator, is_fitted_by=self.is_fitted):
-            # Fit the wrapped estimator
-            self.estimator.fit(X, y, **kwargs)
+            if hasattr(self.estimator, "fit_predict"):
+                labels = self.estimator.fit_predict(X, y, **kwargs)
+            else:
+                self.estimator.fit(X, y, **kwargs)
+                labels = self.estimator.predict(X)
+        else:
+            if hasattr(self.estimator, "predict"):
+                labels = self.estimator.predict(X)
+            else:
+                labels = self.estimator.fit_predict(X, y, **kwargs)
 
         # Get the properties of the dataset
         self.n_samples_ = X.shape[0]
-        self.n_clusters_ = self.estimator.n_clusters
+
+        # Compute the number of available clusters from the estimator
+        if hasattr(self.estimator, "n_clusters"):
+            self.n_clusters_ = self.estimator.n_clusters
+        else:
+            unique_labels = set(labels)
+            n_noise_clusters = 1 if -1 in unique_labels else 0
+            self.n_clusters_ = len(unique_labels) - n_noise_clusters
+
+        # Identify the distance metric to use for silhouette scoring
+        metric = self._identify_silhouette_metric()
 
         # Compute the scores of the cluster
-        labels = self.estimator.predict(X)
-        self.silhouette_score_ = silhouette_score(X, labels)
-        self.silhouette_samples_ = silhouette_samples(X, labels)
+        self.silhouette_score_ = silhouette_score(X, labels, metric=metric)
+        self.silhouette_samples_ = silhouette_samples(X, labels, metric=metric)
 
         # Draw the silhouette figure
         self.draw(labels)
@@ -185,7 +237,6 @@ def draw(self, labels):
         # For each cluster, plot the silhouette scores
         self.y_tick_pos_ = []
         for idx in range(self.n_clusters_):
-
             # Collect silhouette scores for samples in the current cluster .
             values = self.silhouette_samples_[labels == idx]
             values.sort()
@@ -260,6 +311,26 @@ def finalize(self):
         # Show legend (Average Silhouette Score axis)
         self.ax.legend(loc="best")
 
+    def _identify_silhouette_metric(self):
+        """
+        The Silhouette metric must be one of the distance options allowed by
+        metrics.pairwise.pairwise_distances or a callable. This method attempts to
+        discover a valid distance metric from the underlying estimator or returns
+        "euclidean" by default.
+        """
+        if hasattr(self.estimator, "metric"):
+            if callable(self.estimator.metric):
+                return self.estimator.metric
+
+            if self.estimator.metric in _VALID_METRICS:
+                return self.estimator.metric
+
+        if hasattr(self.estimator, "affinity"):
+            if self.estimator.affinity in _VALID_METRICS:
+                return self.estimator.affinity
+
+        return "euclidean"
+
 
 ##########################################################################
 ## Quick Method