Spark Logistic回归用于二元分类：为预测2类应用新的阈值

Question

I am new to both Spark and Python. I used Spark to train a Logistic Regression model with just two classes (0 and 1). I trained it using my train data frame.

This is how my pipeline model was defined:

# Model definition:
lr = LogisticRegression(featuresCol = "lr_features", labelCol = "targetvar")
# Pipeline definition:
lr_pipeline = Pipeline(stages = indexStages + encodeStages +[lr_assembler, lr])
# Fit the logistic regression model:
lrModel = lr_pipeline.fit(train)

Then I ran predictions using my test dataframe:

lr_predictions = lrModel.transform(test)

Now, my lr_predictions has a column 'probability' that looks like a nested list to me. For example, its first cell contains: [1,2,[],[0.88,0.11]]
I assume, it means: the probability for class 1 (which is = 0) is 0.88, the probability for class 2 (which is = 1) is 0.11.

By default (threshold = 0.5) this observation is predicted as 0. However, I found a value (bestThreshold) that maximizes the F-measure (in my case it's 0.21):

fMeasure = lr_summary.fMeasureByThreshold
bestThreshold = fMeasure.orderBy(fMeasure['F-Measure'].desc()).first().threshold

I would like to apply bestThreshold to the 'probability' column and get a new column ('pred_new', for example) that contains the class assignments (0 or 1) based on bestThreshold as opposed to 0.5.

I cannot use the code below, because 'probability' column is too complex:

from pyspark.sql.functions import when
lr_predictions = lr_predictions.withColumn("prob_best", \
              when(lr_predictions["probability"] >= bestThreshold, 1).otherwise(0)

I feel I need to need to map the 'probability' to a new column based on a new threshold. But I am not sure how to do it - given this complex (for me) structure of the 'probability' column.

Thank you so much for your advice!

Answer 1

If lrModel is LogisticRegressionModel :

type(lrModel)
## pyspark.ml.classification.LogisticRegressionModel

You can use internal Java object to set threshold

lrModel._java_obj.setThreshold(bestThreshold)

and transform:

lrModel.transform(data)

You can do the same to modify rawPredictionCol , predictionCol and probabilityCol .

This should become part of the public API in the future (2.3):

lrModel.transform(data, {lrModel.threshold: bestThreshold})

You can also use UDF:

from pyspark.sql.functions import udf, lit

@udf("integer")
def predict(v, threshold):
    return 0 if v[0] >= bestThreshold  else 1

lr_predictions.withColumn(
   "prob_best",
   predict(lr_predictions["probability"], lit(bestThreshold)))

Edit :

With PipelineModel you can try to access LogisticRegressionModel (as in your previous question ) and do the same thing.