Merge pull request #2350 from kernc/rank-learners

[FIX] Rank widget supports Scorer inputs

Author: janezd

Orange/modelling/base.py

@@ -1,6 +1,7 @@
 import numpy as np
 
 from Orange.base import Learner, Model, SklLearner
+from Orange.data import Table, Domain
 
 
 class Fitter(Learner):
@@ -31,10 +32,7 @@ def __init__(self, preprocessors=None, **kwargs):
         self.__learners = {self.CLASSIFICATION: None, self.REGRESSION: None}
 
     def _fit_model(self, data):
-        if data.domain.has_discrete_class:
-            learner = self.get_learner(self.CLASSIFICATION)
-        else:
-            learner = self.get_learner(self.REGRESSION)
+        learner = self.get_learner(data)
 
         if type(self).fit is Learner.fit:
             return learner.fit_storage(data)
@@ -43,20 +41,34 @@ def _fit_model(self, data):
             return learner.fit(X, Y, W)
 
     def preprocess(self, data):
-        if data.domain.has_discrete_class:
-            return self.get_learner(self.CLASSIFICATION).preprocess(data)
-        else:
-            return self.get_learner(self.REGRESSION).preprocess(data)
+        return self.get_learner(data).preprocess(data)
 
     def get_learner(self, problem_type):
         """Get the learner for a given problem type.
 
+        Parameters
+        ----------
+        problem_type: str or Table or Domain
+            If str, one of ``'classification'`` or ``'regression'``. If Table
+            or Domain, the type is inferred from Domain's first class variable.
+
         Returns
         -------
         Learner
             The appropriate learner for the given problem type.
 
+        Raises
+        ------
+        TypeError
+            When (inferred) problem type not one of ``'classification'``
+            or ``'regression'``.
         """
+        if isinstance(problem_type, Table):
+            problem_type = problem_type.domain
+        if isinstance(problem_type, Domain):
+            problem_type = (self.CLASSIFICATION if problem_type.has_discrete_class else
+                            self.REGRESSION if problem_type.has_continuous_class else
+                            None)
         # Prevent trying to access the learner when problem type is None
         if problem_type not in self.__fits__:
             raise TypeError("No learner to handle '{}'".format(problem_type))
@@ -112,8 +124,5 @@ class SklFitter(Fitter):
     def _fit_model(self, data):
         model = super()._fit_model(data)
         model.used_vals = [np.unique(y) for y in data.Y[:, None].T]
-        if data.domain.has_discrete_class:
-            model.params = self.get_params(self.CLASSIFICATION)
-        else:
-            model.params = self.get_params(self.REGRESSION)
+        model.params = self.get_params(data)
         return model

Orange/modelling/linear.py

@@ -1,11 +1,24 @@
+import numpy as np
+
 from Orange.classification.sgd import SGDClassificationLearner
+from Orange.data import Variable
 from Orange.modelling import SklFitter
+from Orange.preprocess.score import LearnerScorer
 from Orange.regression import SGDRegressionLearner
 
 __all__ = ['SGDLearner']
 
 
-class SGDLearner(SklFitter):
+class _FeatureScorerMixin(LearnerScorer):
+    feature_type = Variable
+    class_type = Variable
+
+    def score(self, data):
+        model = self.get_learner(data)(data)
+        return np.atleast_2d(np.abs(model.skl_model.coef_)).mean(0)
+
+
+class SGDLearner(SklFitter, _FeatureScorerMixin):
     name = 'sgd'
 
     __fits__ = {'classification': SGDClassificationLearner,

Orange/modelling/randomforest.py

@@ -1,12 +1,23 @@
 from Orange.base import RandomForestModel
 from Orange.classification import RandomForestLearner as RFClassification
+from Orange.data import Variable
 from Orange.modelling import SklFitter
+from Orange.preprocess.score import LearnerScorer
 from Orange.regression import RandomForestRegressionLearner as RFRegression
 
 __all__ = ['RandomForestLearner']
 
 
-class RandomForestLearner(SklFitter):
+class _FeatureScorerMixin(LearnerScorer):
+    feature_type = Variable
+    class_type = Variable
+
+    def score(self, data):
+        model = self.get_learner(data)(data)
+        return model.skl_model.feature_importances_
+
+
+class RandomForestLearner(SklFitter, _FeatureScorerMixin):
     name = 'random forest'
 
     __fits__ = {'classification': RFClassification,

Orange/preprocess/score.py

@@ -154,11 +154,10 @@ def score(self, data):
         raise NotImplementedError
 
     def score_data(self, data, feature=None):
-        scores = self.score(data)
 
         def average_scores(scores):
             scores_grouped = defaultdict(list)
-            for attr, score in zip(self.domain.attributes, scores):
+            for attr, score in zip(model_domain.attributes, scores):
                 # Go up the chain of preprocessors to obtain the original variable
                 while getattr(attr, 'compute_value', False):
                     attr = getattr(attr.compute_value, 'variable', attr)
@@ -167,8 +166,14 @@ def average_scores(scores):
                     if attr in scores_grouped else 0
                     for attr in data.domain.attributes]
 
-        scores = np.atleast_2d(scores)
-        if data.domain != self.domain:
+        scores = np.atleast_2d(self.score(data))
+
+        from Orange.modelling import Fitter  # Avoid recursive import
+        model_domain = (self.get_learner(data).domain
+                        if isinstance(self, Fitter) else
+                        self.domain)
+
+        if data.domain != model_domain:
             scores = np.array([average_scores(row) for row in scores])
 
         return scores[:, data.domain.attributes.index(feature)] \

Orange/widgets/data/tests/test_owrank.py

@@ -1,13 +1,16 @@
 import numpy as np
 
 from Orange.data import Table, Domain, ContinuousVariable, DiscreteVariable
+from Orange.modelling import RandomForestLearner, SGDLearner
 from Orange.preprocess.score import Scorer
 from Orange.classification import LogisticRegressionLearner
 from Orange.regression import LinearRegressionLearner
 from Orange.projection import PCA
 from Orange.widgets.data.owrank import OWRank
 from Orange.widgets.tests.base import WidgetTest
 
+from AnyQt.QtCore import Qt
+
 
 class TestOWRank(WidgetTest):
     def setUp(self):
@@ -39,6 +42,31 @@ def test_input_scorer(self):
         self.assertEqual(self.log_reg, value.score)
         self.assertIsInstance(value.score, Scorer)
 
+    def test_input_scorer_fitter(self):
+        heart_disease = Table('heart_disease')
+        self.assertEqual(self.widget.learners, {})
+
+        for fitter, name in ((RandomForestLearner(), 'random forest'),
+                             (SGDLearner(), 'sgd')):
+            with self.subTest(fitter=fitter):
+                self.send_signal("Scorer", fitter, 1)
+
+                for data, model in ((self.housing, self.widget.contRanksModel),
+                                    (heart_disease, self.widget.discRanksModel)):
+                    with self.subTest(data=data.name):
+                        self.send_signal('Data', data)
+                        scores = [model.data(model.index(row, model.columnCount() - 1))
+                                  for row in range(model.rowCount())]
+                        self.assertEqual(len(scores), len(data.domain.attributes))
+                        self.assertFalse(np.isnan(scores).any())
+
+                        last_column = model.headerData(
+                            model.columnCount() - 1, Qt.Horizontal).lower()
+                        self.assertIn(name, last_column)
+
+                self.send_signal("Scorer", None, 1)
+                self.assertEqual(self.widget.learners, {})
+
     def test_input_scorer_disconnect(self):
         """Check widget's scorer after disconnecting scorer on the input"""
         self.send_signal(self.widget.Inputs.scorer, self.log_reg, 1)

doc/visual-programming/source/widgets/data/rank.rst

@@ -14,6 +14,11 @@ Signals
 
    An input data set.
 
+- **Scorer**  (multiple)
+
+  Models that implement the feature scoring interface, such as linear /
+  logistic regression, random forest, stochastic gradient descent, etc.
+
 **Outputs**:
 
 -  **Reduced Data**
@@ -47,6 +52,12 @@ Scoring methods
 6. `ReliefF <https://en.wikipedia.org/wiki/Relief_(feature_selection)>`_: the ability of an attribute to distinguish between classes on similar data instances
 7. `FCBF (Fast Correlation Based Filter) <https://www.aaai.org/Papers/ICML/2003/ICML03-111.pdf>`_: entropy-based measure, which also identifies redundancy due to pairwise correlations between features
 
+Additionally, you can connect certain learners that enable scoring the features
+according to how important they are in models that the learners build (e.g.
+:ref:`Linear <model.lr>` / :ref:`Logistic Regression <model.logit>`,
+:ref:`Random Forest <model.rf>`, :ref:`SGD <model.sgd>`, …).
+
+
 Example: Attribute Ranking and Selection
 ----------------------------------------
 

doc/visual-programming/source/widgets/model/linearregression.rst

@@ -1,3 +1,5 @@
+.. _model.lr:
+
 Linear Regression
 =================
 

doc/visual-programming/source/widgets/model/logisticregression.rst

@@ -1,3 +1,5 @@
+.. _model.logit:
+
 Logistic Regression
 ===================
 

doc/visual-programming/source/widgets/model/randomforest.rst

@@ -1,3 +1,5 @@
+.. _model.rf:
+
 Random Forest
 =============
 

doc/visual-programming/source/widgets/model/stochasticgradient.rst

@@ -1,3 +1,5 @@
+.. _model.sgd:
+
 Stochastic Gradient Descent
 ===========================