murp?

Author: holdenk

mllib/src/main/scala/org/apache/spark/ml/regression/LinearRegression.scala

@@ -41,7 +41,8 @@ import org.apache.spark.util.StatCounter
  * Params for linear regression.
  */
 private[regression] trait LinearRegressionParams extends PredictorParams
-  with HasRegParam with HasElasticNetParam with HasMaxIter with HasTol
+    with HasRegParam with HasElasticNetParam with HasMaxIter with HasTol with
+    HasIntercept
 
 /**
  * :: Experimental ::
@@ -121,8 +122,9 @@ class LinearRegression(override val uid: String)
       })
 
     val numFeatures = summarizer.mean.size
-    val yMean = statCounter.mean
-    val yStd = math.sqrt(statCounter.variance)
+    val yMean = if (hasIntercept) statCounter.mean else 0.0
+    val yStd = if (hasIntercept) math.sqrt(statCounter.variance) else
+      // look at glmnet6.m L761 maaaybe that has info
 
     // If the yStd is zero, then the intercept is yMean with zero weights;
     // as a result, training is not needed.
@@ -180,6 +182,7 @@ class LinearRegression(override val uid: String)
     // The intercept in R's GLMNET is computed using closed form after the coefficients are
     // converged. See the following discussion for detail.
     // http://stats.stackexchange.com/questions/13617/how-is-the-intercept-computed-in-glmnet
+    // Also see the scikit learn impl at https://github.com/scikit-learn/scikit-learn/blob/master/sklearn/linear_model/base.py
     val intercept = yMean - dot(weights, Vectors.dense(featuresMean))
     if (handlePersistence) instances.unpersist()