Description
SparkR::glm output different features compared with R glm when fit w/o intercept and having string/category features. Take the following example, SparkR output three features compared with four features for native R.
SparkR::glm
training <- suppressWarnings(createDataFrame(sqlContext, iris))
model <- glm(Sepal_Width ~ Sepal_Length + Species - 1, data = training)
summary(model)Coefficients:
Estimate Std. Error t value Pr(>|t|)
Sepal_Length 0.67468 0.0093013 72.536 0
Species_versicolor -1.2349 0.07269 -16.989 0
Species_virginica -1.4708 0.077397 -19.003 0
stats::glm
summary(glm(Sepal.Width ~ Sepal.Length + Species - 1, data = iris))
Coefficients:
Estimate Std. Error t value Pr(>|t|)
Sepal.Length 0.3499 0.0463 7.557 4.19e-12 ***
Speciessetosa 1.6765 0.2354 7.123 4.46e-11 ***
Speciesversicolor 0.6931 0.2779 2.494 0.0137 *
Speciesvirginica 0.6690 0.3078 2.174 0.0313 *
The encoder for string/category feature is different. R did not drop any category but SparkR drop the last one.
I searched online and test some other cases, found when we fit R glm model(or other models powered by R formula) w/o intercept on a dataset including string/category features, one of the categories in the first category feature is being used as reference category, we will not drop any category for that feature.
I think we should keep consistent semantics between Spark RFormula and R formula.
cc mengxr
Attachments
Issue Links
- Is contained by
-
SPARK-15540 RFormula and R feature processing improvement umbrella
-
- Resolved
-
- is related to
-
SPARK-14659 OneHotEncoder support drop first category alphabetically in the encoded vector
-
- Resolved
-
- links to
