Skip to content

[SPARK-18701][ML] Fix Poisson GLM failure due to wrong initialization #16131

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 5 commits into from
Closed

[SPARK-18701][ML] Fix Poisson GLM failure due to wrong initialization #16131

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
784cb09
Change initial value in Poisson GLM to avoid numerical issue
actuaryzhang Dec 4, 2016
56c4779
Update Poisson GLM test (for incorrect initialization)
actuaryzhang Dec 4, 2016
a24966a
Set lower bound in Poisson GLM initialization
actuaryzhang Dec 5, 2016
b162970
Update comment and remove unnecessary printing in test
actuaryzhang Dec 6, 2016
271d315
Add reference to R's Poisson GLM initialization
actuaryzhang Dec 6, 2016
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 5 additions & 1 deletion mllib/src/main/scala/org/apache/spark/ml/regression/GeneralizedLinearRegression.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -505,7 +505,11 @@ object GeneralizedLinearRegression extends DefaultParamsReadable[GeneralizedLine
override def initialize(y: Double, weight: Double): Double = {
require(y >= 0.0, "The response variable of Poisson family " +
s"should be non-negative, but got $y")
y
/*
Force Poisson mean > 0 to avoid numerical instability in IRLS.
R uses y + 0.1 for initialization. See poisson()$initialize.
*/
math.max(y, 0.1)
}

override def variance(mu: Double): Double = mu
Expand Down
21 changes: 12 additions & 9 deletions mllib/src/test/scala/org/apache/spark/ml/regression/GeneralizedLinearRegressionSuite.scala
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -89,11 +89,14 @@ class GeneralizedLinearRegressionSuite
xVariance = Array(0.7, 1.2), nPoints = 10000, seed, noiseLevel = 0.01,
family = "poisson", link = "log").toDF()

datasetPoissonLogWithZero = generateGeneralizedLinearRegressionInput(
intercept = -1.5, coefficients = Array(0.22, 0.06), xMean = Array(2.9, 10.5),
xVariance = Array(0.7, 1.2), nPoints = 100, seed, noiseLevel = 0.01,
family = "poisson", link = "log")
.map{x => LabeledPoint(if (x.label < 0.7) 0.0 else x.label, x.features)}.toDF()
datasetPoissonLogWithZero = Seq(
LabeledPoint(0.0, Vectors.dense(18, 1.0)),
LabeledPoint(1.0, Vectors.dense(12, 0.0)),
LabeledPoint(0.0, Vectors.dense(15, 0.0)),
LabeledPoint(0.0, Vectors.dense(13, 2.0)),
LabeledPoint(0.0, Vectors.dense(15, 1.0)),
LabeledPoint(1.0, Vectors.dense(16, 1.0))
).toDF()

datasetPoissonIdentity = generateGeneralizedLinearRegressionInput(
intercept = 2.5, coefficients = Array(2.2, 0.6), xMean = Array(2.9, 10.5),
Expand Down Expand Up @@ -480,12 +483,12 @@ class GeneralizedLinearRegressionSuite
model <- glm(formula, family="poisson", data=data)
print(as.vector(coef(model)))
}
[1] 0.4272661 -0.1565423
[1] -3.6911354 0.6214301 0.1295814
[1] -0.0457441 -0.6833928
[1] 1.8121235 -0.1747493 -0.5815417
*/
val expected = Seq(
Vectors.dense(0.0, 0.4272661, -0.1565423),
Vectors.dense(-3.6911354, 0.6214301, 0.1295814))
Vectors.dense(0.0, -0.0457441, -0.6833928),
Vectors.dense(1.8121235, -0.1747493, -0.5815417))

import GeneralizedLinearRegression._

Expand Down