From eb00378f0eed6afbf328ae6cd541cc202d14c1f0 Mon Sep 17 00:00:00 2001 From: WeichenXu Date: Fri, 21 Apr 2017 17:58:13 +0000 Subject: [SPARK-20423][ML] fix MLOR coeffs centering when reg == 0 ## What changes were proposed in this pull request? When reg == 0, MLOR has multiple solutions and we need to centralize the coeffs to get identical result. BUT current implementation centralize the `coefficientMatrix` by the global coeffs means. In fact the `coefficientMatrix` should be centralized on each feature index itself. Because, according to the MLOR probability distribution function, it can be proven easily that: suppose `{ w0, w1, .. w(K-1) }` make up the `coefficientMatrix`, then `{ w0 + c, w1 + c, ... w(K - 1) + c}` will also be the equivalent solution. `c` is an arbitrary vector of `numFeatures` dimension. reference https://core.ac.uk/download/pdf/6287975.pdf So that we need to centralize the `coefficientMatrix` on each feature dimension separately. **We can also confirm this through R library `glmnet`, that MLOR in `glmnet` always generate coefficients result that the sum of each dimension is all `zero`, when reg == 0.** ## How was this patch tested? Tests added. Author: WeichenXu Closes #17706 from WeichenXu123/mlor_center. --- .../apache/spark/ml/classification/LogisticRegression.scala | 11 ++++++++--- .../spark/ml/classification/LogisticRegressionSuite.scala | 6 ++++++ 2 files changed, 14 insertions(+), 3 deletions(-) diff --git a/mllib/src/main/scala/org/apache/spark/ml/classification/LogisticRegression.scala b/mllib/src/main/scala/org/apache/spark/ml/classification/LogisticRegression.scala index 965ce3d6f2..bc8154692e 100644 --- a/mllib/src/main/scala/org/apache/spark/ml/classification/LogisticRegression.scala +++ b/mllib/src/main/scala/org/apache/spark/ml/classification/LogisticRegression.scala @@ -609,9 +609,14 @@ class LogisticRegression @Since("1.2.0") ( Friedman, et al. "Regularization Paths for Generalized Linear Models via Coordinate Descent," https://core.ac.uk/download/files/153/6287975.pdf */ - val denseValues = denseCoefficientMatrix.values - val coefficientMean = denseValues.sum / denseValues.length - denseCoefficientMatrix.update(_ - coefficientMean) + val centers = Array.fill(numFeatures)(0.0) + denseCoefficientMatrix.foreachActive { case (i, j, v) => + centers(j) += v + } + centers.transform(_ / numCoefficientSets) + denseCoefficientMatrix.foreachActive { case (i, j, v) => + denseCoefficientMatrix.update(i, j, v - centers(j)) + } } // center the intercepts when using multinomial algorithm diff --git a/mllib/src/test/scala/org/apache/spark/ml/classification/LogisticRegressionSuite.scala b/mllib/src/test/scala/org/apache/spark/ml/classification/LogisticRegressionSuite.scala index c858b9bbfc..83f575e838 100644 --- a/mllib/src/test/scala/org/apache/spark/ml/classification/LogisticRegressionSuite.scala +++ b/mllib/src/test/scala/org/apache/spark/ml/classification/LogisticRegressionSuite.scala @@ -1139,6 +1139,9 @@ class LogisticRegressionSuite 0.10095851, -0.85897154, 0.08392798, 0.07904499), isTransposed = true) val interceptsR = Vectors.dense(-2.10320093, 0.3394473, 1.76375361) + model1.coefficientMatrix.colIter.foreach(v => assert(v.toArray.sum ~== 0.0 absTol eps)) + model2.coefficientMatrix.colIter.foreach(v => assert(v.toArray.sum ~== 0.0 absTol eps)) + assert(model1.coefficientMatrix ~== coefficientsR relTol 0.05) assert(model1.coefficientMatrix.toArray.sum ~== 0.0 absTol eps) assert(model1.interceptVector ~== interceptsR relTol 0.05) @@ -1204,6 +1207,9 @@ class LogisticRegressionSuite -0.3180040, 0.9679074, -0.2252219, -0.4319914, 0.2452411, -0.6046524, 0.1050710, 0.1180180), isTransposed = true) + model1.coefficientMatrix.colIter.foreach(v => assert(v.toArray.sum ~== 0.0 absTol eps)) + model2.coefficientMatrix.colIter.foreach(v => assert(v.toArray.sum ~== 0.0 absTol eps)) + assert(model1.coefficientMatrix ~== coefficientsR relTol 0.05) assert(model1.coefficientMatrix.toArray.sum ~== 0.0 absTol eps) assert(model1.interceptVector.toArray === Array.fill(3)(0.0)) -- cgit v1.2.3