Stabilize SDCA logistic regression test (#7677)

The strict SdcaLogisticRegression quality assertions now use a single training thread, making their results deterministic. A separate four-thread test retains correctness coverage of SDCA's nondeterministic parallel path by preserving the original AUC requirement and verifying that its LogLoss outperforms a featureless Prior baseline.

My theory is that the macOS arm64 runners have more threads than the Windows/Linux runners which leads to more non-determinism, triggering the flaky failure.

Author: matouskozak

test/Microsoft.ML.Tests/TrainerEstimators/SdcaTests.cs

@@ -72,7 +72,14 @@ public void SdcaLogisticRegression()
 
             // Step 2: Create a binary classifier.
             // We set the "Label" column as the label of the dataset, and the "Features" column as the features column.
-            var pipeline = mlContext.BinaryClassification.Trainers.SdcaLogisticRegression(labelColumnName: "Label", featureColumnName: "Features", l2Regularization: 0.001f);
+            var pipeline = mlContext.BinaryClassification.Trainers.SdcaLogisticRegression(
+                new SdcaLogisticRegressionBinaryTrainer.Options
+                {
+                    LabelColumnName = "Label",
+                    FeatureColumnName = "Features",
+                    L2Regularization = 0.001f,
+                    NumberOfThreads = 1
+                });
 
             // Step 3: Train the pipeline created.
             var model = pipeline.Fit(data);
@@ -97,6 +104,44 @@ public void SdcaLogisticRegression()
             Assert.InRange(first.Probability, 0.8, 1);
         }
 
+        [Fact]
+        public void SdcaLogisticRegressionMultithreaded()
+        {
+            // Keep coverage of the nondeterministic parallel path while the quality test above remains deterministic.
+            // See https://github.com/dotnet/machinelearning/blob/240a849becda954f3e17d39f7606328be3a7f0de/src/Microsoft.ML.StandardTrainers/Standard/SdcaBinary.cs#L180-L188.
+
+            // Generate C# objects as training examples.
+            var rawData = SamplesUtils.DatasetUtils.GenerateBinaryLabelFloatFeatureVectorFloatWeightSamples(100);
+
+            // Create a new context for ML.NET operations.
+            var mlContext = new MLContext(1);
+
+            // Step 1: Read and cache the data as an IDataView.
+            var data = mlContext.Data.Cache(mlContext.Data.LoadFromEnumerable(rawData));
+
+            // Step 2: Create a binary classifier that exercises the parallel training path.
+            var pipeline = mlContext.BinaryClassification.Trainers.SdcaLogisticRegression(
+                new SdcaLogisticRegressionBinaryTrainer.Options
+                {
+                    LabelColumnName = "Label",
+                    FeatureColumnName = "Features",
+                    L2Regularization = 0.001f,
+                    NumberOfThreads = 4
+                });
+
+            // Step 3: Train and evaluate the classifier on the training data.
+            var transformedData = pipeline.Fit(data).Transform(data);
+            var metrics = mlContext.BinaryClassification.Evaluate(transformedData);
+
+            // Step 4: Evaluate a featureless classifier as a quality baseline.
+            var priorModel = mlContext.BinaryClassification.Trainers.Prior().Fit(data);
+            var priorMetrics = mlContext.BinaryClassification.Evaluate(priorModel.Transform(data));
+
+            // Verify the multithreaded classifier learns the signal in the generated data.
+            Assert.InRange(metrics.AreaUnderRocCurve, 0.9, 1);
+            Assert.True(metrics.LogLoss < priorMetrics.LogLoss);
+        }
+
         [Fact]
         public void SdcaLogisticRegressionWithWeight()
         {