[Spark-21854] Added LogisticRegressionTrainingSummary for MultinomialLogisticRegression in Python API

python/pyspark/ml/classification.py

@@ -529,8 +529,11 @@ def summary(self):
         """
         if self.hasSummary:
             java_blrt_summary = self._call_java("summary")
-            # Note: Once multiclass is added, update this to return correct summary
-            return BinaryLogisticRegressionTrainingSummary(java_blrt_summary)
+            if (self.numClasses == 2):
+                java_blrt_binarysummary = self._call_java("binarySummary")
+                return BinaryLogisticRegressionTrainingSummary(java_blrt_binarysummary)
+            else:
+                return LogisticRegressionTrainingSummary(java_blrt_summary)
         else:
             raise RuntimeError("No training summary available for this %s" %
                                self.__class__.__name__)
@@ -585,6 +588,14 @@ def probabilityCol(self):
         """
         return self._call_java("probabilityCol")
 
+    @property
+    @since("2.3.0")
+    def predictionCol(self):
+        """
+        Field in "predictions" which gives the prediction of each class.
+        """
+        return self._call_java("predictionCol")
+
     @property
     @since("2.0.0")
     def labelCol(self):
@@ -603,6 +614,112 @@ def featuresCol(self):
         """
         return self._call_java("featuresCol")
 
+    @property
+    @since("2.3.0")
+    def labels(self):
+        """
+        Returns the sequence of labels in ascending order. This order matches the order used
+        in metrics which are specified as arrays over labels, e.g., truePositiveRateByLabel.
+
+        Note: In most cases, it will be values {0.0, 1.0, ..., numClasses-1}, However, if the
+        training set is missing a label, then all of the arrays over labels
+        (e.g., from truePositiveRateByLabel) will be of length numClasses-1 instead of the
+        expected numClasses.
+        """
+        return self._call_java("labels")
+
+    @property
+    @since("2.3.0")
+    def truePositiveRateByLabel(self):
+        """
+        Returns true positive rate for each label (category).
+        """
+        return self._call_java("truePositiveRateByLabel")
+
+    @property
+    @since("2.3.0")
+    def falsePositiveRateByLabel(self):
+        """
+        Returns false positive rate for each label (category).
+        """
+        return self._call_java("falsePositiveRateByLabel")
+
+    @property
+    @since("2.3.0")
+    def precisionByLabel(self):
+        """
+        Returns precision for each label (category).
+        """
+        return self._call_java("precisionByLabel")
+
+    @property
+    @since("2.3.0")
+    def recallByLabel(self):
+        """
+        Returns recall for each label (category).
+        """
+        return self._call_java("recallByLabel")
+
+    @property
+    @since("2.3.0")
+    def fMeasureByLabel(self, beta=1.0):
+        """
+        Returns f-measure for each label (category).
+        """
+        return self._call_java("fMeasureByLabel", beta)
+
+    @property
+    @since("2.3.0")
+    def accuracy(self):
+        """
+        Returns accuracy.
+        (equals to the total number of correctly classified instances
+        out of the total number of instances.)
+        """
+        return self._call_java("accuracy")
+
+    @property
+    @since("2.3.0")
+    def weightedTruePositiveRate(self):
+        """
+        Returns weighted true positive rate.
+        (equals to precision, recall and f-measure)
+        """
+        return self._call_java("weightedTruePositiveRate")
+
+    @property
+    @since("2.3.0")
+    def weightedFalsePositiveRate(self):
+        """
+        Returns weighted false positive rate.
+        """
+        return self._call_java("weightedFalsePositiveRate")
+
+    @property
+    @since("2.3.0")
+    def weightedRecall(self):
+        """
+        Returns weighted averaged recall.
+        (equals to precision, recall and f-measure)
+        """
+        return self._call_java("weightedRecall")
+
+    @property
+    @since("2.3.0")
+    def weightedPrecision(self):
+        """
+        Returns weighted averaged precision.
+        """
+        return self._call_java("weightedPrecision")
+
+    @property
+    @since("2.3.0")
+    def weightedFMeasure(self, beta=1.0):
+        """
+        Returns weighted averaged f-measure.
+        """
+        return self._call_java("weightedFMeasure", beta)
+
 
 @inherit_doc
 class LogisticRegressionTrainingSummary(LogisticRegressionSummary):

python/pyspark/ml/tests.py

@@ -1478,6 +1478,40 @@ def test_logistic_regression_summary(self):
         sameSummary = model.evaluate(df)
         self.assertAlmostEqual(sameSummary.areaUnderROC, s.areaUnderROC)
 
+    def test_multiclass_logistic_regression_summary(self):
+        df = self.spark.createDataFrame([(1.0, 2.0, Vectors.dense(1.0)),
+                                         (0.0, 2.0, Vectors.sparse(1, [], [])),
+                                         (2.0, 2.0, Vectors.dense(2.0)),
+                                         (2.0, 2.0, Vectors.dense(1.9))],
+                                        ["label", "weight", "features"])
+        lr = LogisticRegression(maxIter=5, regParam=0.01, weightCol="weight", fitIntercept=False)
+        model = lr.fit(df)
+        self.assertTrue(model.hasSummary)
+        s = model.summary
+        # test that api is callable and returns expected types
+        self.assertTrue(isinstance(s.predictions, DataFrame))
+        self.assertEqual(s.probabilityCol, "probability")
+        self.assertEqual(s.labelCol, "label")
+        self.assertEqual(s.featuresCol, "features")
+        self.assertEqual(s.predictionCol, "prediction")
+        objHist = s.objectiveHistory
+        self.assertTrue(isinstance(objHist, list) and isinstance(objHist[0], float))
+        self.assertGreater(s.totalIterations, 0)
+        self.assertTrue(isinstance(s.labels, list))
+        self.assertTrue(isinstance(s.truePositiveRateByLabel, list))
+        self.assertTrue(isinstance(s.falsePositiveRateByLabel, list))
+        self.assertTrue(isinstance(s.precisionByLabel, list))
+        self.assertTrue(isinstance(s.recallByLabel, list))
+        self.assertTrue(isinstance(s.fMeasureByLabel, list))
+        self.assertAlmostEqual(s.accuracy, 0.75, 2)
+        self.assertAlmostEqual(s.weightedTruePositiveRate, 0.75, 2)
+        self.assertAlmostEqual(s.weightedFalsePositiveRate, 0.25, 2)
+        self.assertAlmostEqual(s.weightedRecall, 0.75, 2)
+        self.assertAlmostEqual(s.weightedPrecision, 0.583, 2)
+        self.assertAlmostEqual(s.weightedFMeasure, 0.65, 2)
+        # test evaluation (with training dataset) produces a summary with same values
+        # one check is enough to verify a summary is returned, Scala version runs full test
+
     def test_gaussian_mixture_summary(self):
         data = [(Vectors.dense(1.0),), (Vectors.dense(5.0),), (Vectors.dense(10.0),),
                 (Vectors.sparse(1, [], []),)]