How to convert ArrayType to DenseVector in PySpark

问题:

I'm getting the following error trying to build a ML Pipeline:

pyspark.sql.utils.IllegalArgumentException: 'requirement failed: Column features must be of type org.apache.spark.ml.linalg.VectorUDT@3bfc3ba7 but was actually ArrayType(DoubleType,true).'

My features column contains an array of floating point values. It sounds like I need to convert those to some type of vector (it's not sparse, so a DenseVector?). Is there a way to do this directly on the DataFrame or do I need to convert to an RDD?

回答1:

You can use UDF:

udf(lambda vs: Vectors.dense(vs), VectorUDT())

In Spark < 2.0 import:

from pyspark.mllib.linalg import Vectors, VectorUDT

In Spark 2.0+ import:

from pyspark.ml.linalg import Vectors, VectorUDT

Please note that these classes are not compatible despite identical implementation.

It is also possible to extract individual features and assemble with VectorAssembler. Assuming input column is called features:

from pyspark.ml.feature import VectorAssembler

n = ... # Size of features

assembler = VectorAssembler(
    inputCols=["features[{0}]".format(i) for i in range(n)], 
    outputCol="features_vector")

assembler.transform(df.select(
    "*", *(df["features"].getItem(i) for i in range(n))
))