Update predict_proba of classifier by diffusion

sknetwork-team · May 22, 2023 · 31dd1a5 · 31dd1a5
1 parent d64011d
commit 31dd1a5
Show file tree

Hide file tree

Showing 5 changed files with 20 additions and 71 deletions.
diff --git a/docs/reference/linalg.rst b/docs/reference/linalg.rst
@@ -5,6 +5,13 @@ Linear algebra
 
 Tools of linear algebra.
 
+Normalization
+-------------
+
+.. autofunction:: sknetwork.linalg.normalize
+
+.. autofunction:: sknetwork.linalg.diagonal_pseudo_inverse
+
 Sparse + Low Rank
 -----------------
 
@@ -17,10 +24,3 @@ Solvers
 
 .. _lanczossvd:
 .. autoclass:: sknetwork.linalg.LanczosSVD
-
-Normalization
--------------
-
-.. autofunction:: sknetwork.linalg.normalize
-
-.. autofunction:: sknetwork.linalg.diagonal_pseudo_inverse
diff --git a/docs/tutorials/classification/diffusion.ipynb b/docs/tutorials/classification/diffusion.ipynb
@@ -209,49 +209,6 @@
     "SVG(image)"
    ]
   },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "pycharm": {
-     "name": "#%%\n"
-    }
-   },
-   "outputs": [],
-   "source": [
-    "# threshold\n",
-    "diffusion = DiffusionClassifier(threshold=0.1)\n",
-    "labels_pred = diffusion.fit_predict(adjacency, labels)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "pycharm": {
-     "name": "#%%\n"
-    }
-   },
-   "outputs": [],
-   "source": [
-    "# unclassified node\n",
-    "np.flatnonzero(labels_pred == -1)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {
-    "pycharm": {
-     "name": "#%%\n"
-    }
-   },
-   "outputs": [],
-   "source": [
-    "image = svg_graph(adjacency, position, labels=labels_pred, seeds=labels)\n",
-    "SVG(image)"
-   ]
-  },
   {
    "cell_type": "markdown",
    "metadata": {

diff --git a/sknetwork/classification/diffusion.py b/sknetwork/classification/diffusion.py
@@ -28,9 +28,9 @@ class DiffusionClassifier(BaseClassifier):
         Number of iterations of the diffusion (discrete time).
     centering : bool
         If ``True``, center the temperature of each label to its mean before classification (default).
-    threshold : float
-        Minimum difference of temperatures between the 2 top labels to classify a node (default = 0).
-        If the difference of temperatures does not exceed this threshold, return -1 for this node (no label).
+    scale : float
+        Multiplicative factor applied to tempreatures before softmax (default = 5).
+        Used only when centering is ``True``.
 
     Attributes
     ----------
@@ -60,15 +60,15 @@ class DiffusionClassifier(BaseClassifier):
     Zhu, X., Lafferty, J., & Rosenfeld, R. (2005). `Semi-supervised learning with graphs`
     (Doctoral dissertation, Carnegie Mellon University, language technologies institute, school of computer science).
     """
-    def __init__(self, n_iter: int = 10, centering: bool = True, threshold: float = 0):
+    def __init__(self, n_iter: int = 10, centering: bool = True, scale: float = 5):
         super(DiffusionClassifier, self).__init__()
 
         if n_iter <= 0:
             raise ValueError('The number of iterations must be positive.')
         else:
             self.n_iter = n_iter
         self.centering = centering
-        self.threshold = threshold
+        self.scale = scale
 
     def fit(self, input_matrix: Union[sparse.csr_matrix, np.ndarray],
             labels: Optional[Union[dict, np.ndarray]] = None, labels_row: Optional[Union[dict, np.ndarray]] = None,
@@ -107,25 +107,17 @@ def fit(self, input_matrix: Union[sparse.csr_matrix, np.ndarray],
             temperatures[labels >= 0] = temperatures_seeds
         if self.centering:
             temperatures -= temperatures.mean(axis=0)
-            temperatures = np.maximum(temperatures, 0)
-
         labels_ = temperatures.argmax(axis=1)
 
+        # softmax
+        if self.centering:
+            temperatures = np.exp(self.scale * temperatures)
+
         # set label -1 to nodes not reached by diffusion
         distances = get_distances(adjacency, source=np.flatnonzero(labels >= 0))
         labels_[distances < 0] = -1
         temperatures[distances < 0] = 0
 
-        # set label -1 to nodes with low confidence
-        if self.threshold >= 0:
-            n_labels = temperatures.shape[1]
-            if n_labels > 2:
-                top_temperatures = np.partition(-temperatures, 2, axis=1)[:, :2]
-            else:
-                top_temperatures = temperatures
-            gap = np.abs(top_temperatures[:, 0] - top_temperatures[:, 1])
-            labels_[gap <= self.threshold] = -1
-
         self.labels_ = labels_
         self.probs_ = sparse.csr_matrix(normalize(temperatures))
         self._split_vars(input_matrix.shape)

diff --git a/sknetwork/classification/tests/test_diffusion.py b/sknetwork/classification/tests/test_diffusion.py
@@ -23,9 +23,9 @@ def test_graph(self):
         self.assertTrue(len(algo.labels_) == n_nodes)
         with self.assertRaises(ValueError):
             DiffusionClassifier(n_iter=0)
-        algo = DiffusionClassifier(centering=False, threshold=1)
-        algo.fit(adjacency, labels=labels)
-        self.assertTrue(max(algo.labels_) == -1)
+        algo = DiffusionClassifier(centering=True, scale=10)
+        probs = algo.fit_predict_proba(adjacency, labels=labels)[:, 1]
+        self.assertTrue(max(probs) > 0.99)
 
     def test_bipartite(self):
         biadjacency = test_bigraph()

diff --git a/sknetwork/linalg/normalization.py b/sknetwork/linalg/normalization.py
@@ -74,7 +74,7 @@ def normalize(matrix: Union[sparse.csr_matrix, np.ndarray, LinearOperator], p=1)
     Returns
     -------
     normalized matrix :
-        Normalized matrix.
+        Normalized matrix (same format as input matrix).
     """
     norms = get_norms(matrix, p)
     diag = diagonal_pseudo_inverse(norms)