Access element of a vector in a Spark DataFrame (Logistic Regression probability vector) [duplicate]

Posted on

Question :

Access element of a vector in a Spark DataFrame (Logistic Regression probability vector) [duplicate]

I trained a LogisticRegression model in PySpark (ML package) and the result of the prediction is a PySpark DataFrame (cv_predictions) (see [1]). The probability column (see [2]) is a vector type (see [3]).


[2]'probability').show(10, False)
|probability                             |
|[0.8615878905395029,0.1384121094604972] |
|[0.2820729747752462,0.7179270252247538] |
|[0.6346585276598942,0.3653414723401058] |
|[0.6346585276598942,0.3653414723401058] |
|[0.637279255218404,0.362720744781596]   |
only showing top 10 rows

 |-- rawPrediction: vector (nullable = true)
 |-- probability: vector (nullable = true)
 |-- prediction: double (nullable = true)

How do I create parse the vector of the PySpark DataFrame, such that I create a new column that just pulls the first element of each probability vector?

This question is similar to, but the solutions in the links below didn’t work/weren’t clear to me:

How to access the values of denseVector in PySpark

How to access element of a VectorUDT column in a Spark DataFrame?

Answer #1:


It seems like there is a bug in spark that prevents you from accessing individual elements in a dense vector during a select statement. Normally you should would be able to access them just like you would a numpy array, but when trying to run the code previously posted, you may get the error pyspark.sql.utils.AnalysisException: "Can't extract value from probability#12;"

So, one way to handle this to avoid this silly bug is to use a udf. Similar to the other question, you can define a udf in the following way:

from pyspark.sql.functions import udf
from pyspark.sql.types import FloatType

firstelement=udf(lambda v:float(v[0]),FloatType())'probability')).show()

Behind the scenes this still accesses the elements of the DenseVector like a numpy array, but it doesn’t throw the same bug as before.

Since this is getting a lot of upvotes, I figured I should strike through the incorrect portion of this answer.

Original answer:
A dense vector is just a wrapper for a numpy array. So you can access the elements in the same way that you would access the elements of a numpy array.

There are several ways to access individual elements of an array in a dataframe. One is to explicitly call the column cv_predictions_prod['probability'] in your select statement. By explicitly calling the column, you can perform operations on that column, like selecting the first element in the array. For example:['probability'][0]).show()

should solve the problem.

Answered By: DavidWayne

Leave a Reply

Your email address will not be published.