scala/spark map returning an array instead of single value

Question

I have a Scala class that is intended to generalize some functionality of linear models--specifically the user should be able to create an instance with an array of coefficients and an array of predictors, and the class pulls the data from a DataFrame, and uses a simple linear model to create predictions on an entire DataFrame as shown below.

I am stuck on the last line... which I expect to generate a column of predicted values. I have attempted a number of approaches (all but one of which are commented out). The code as it is now wont compile b/c of a type mismatch:

[error]  found   : Array[org.apache.spark.sql.Column]
[error]  required: org.apache.spark.sql.Column
[error]       .withColumn("prediction", colMod(preds.map(p => data(p))))
[error]                                               ^

...which I also get in the pred <- preds version... and the foreach version:

[error]  found   : Unit
[error]  required: org.apache.spark.sql.Column
[error]       .withColumn("prediction", colMod(preds.foreach(data(_))))
[error]                                                   ^

Been trying in vain to resolve... would be grateful for any suggestions.

  class LinearModel(coefficients: Array[Double],
                    predictors: Array[String],
                    data: DataFrame) {

    val coefs = coefficients
    val preds = Array.concat(Array("bias"), predictors)
    require(coefs.length == preds.length)

    /**
      * predict: computes linear model predictions as the dot product of the coefficents and the
      * values (X[i] in the model matrix)
      * @param values: the values from a single row of the given variables from model matrix X
      * @param coefs: array of coefficients to be applied to each of the variables in values
      *             (the first coef is assumed to be 1 for the bias/intercept term)
      * @return: the predicted value 
      */
    private def predict(values: Array[Double], coefs: Array[Double]): Unit = {
      (for ((c, v) <- coefs.zip(values)) yield c * v).sum
    }

    /**
      * colMod (udf): passes the values for each relevant value to predict()
      * @param values: an Array of the numerical values of each of the specified predictors for a
      *              given record
      */
    private val colMod = udf((values: Array[Double]) => predict(values, coefs))

    val dfPred = data
      // create the column with the prediction
      .withColumn("prediction", colMod(preds.map(p => data(p))))
      //.withColumn("prediction", colMod(for (pred <- preds) yield data(pred)))
      //.withColumn("prediction", colMod(preds.foreach(data(_))))
      // prev line should = colMod(data(pred1), data(pred2), ..., data(predn))
  }

zero323 · Accepted Answer · 2016-12-24 09:57:06Z

Here is how it be can done properly:

import org.apache.spark.sql.functions.{lit, col}
import org.apache.spark.sql.Column

def predict(coefficients: Seq[Double], predictors: Seq[String], df: DataFrame) = {

  // I assume there is no predictor for bias
  // but you can easily correct for that
  val prediction: Column = predictors.zip(coefficients).map {
    case (p, c) => col(p) * lit(c)
  }.foldLeft(col("bias"))(_ + _)

  df.withColumn("prediction", prediction)
}

Example usage:

val df = Seq((1.0, -1.0, 3.0, 5.0)).toDF("bias", "x1", "x2", "x3")

predict(Seq(2.0, 3.0), Seq("x1", "x3"), df)

with result being:

+----+----+---+---+----------+
|bias|  x1| x2| x3|prediction|
+----+----+---+---+----------+
| 1.0|-1.0|3.0|5.0|      14.0|
+----+----+---+---+----------+

Regarding your code you've made a number of mistakes:

Array[_] is not a valid external type for ArrayType column. Valid external representation is Seq[_] so the argument of function you pass to udf should be Seq[Double].
Function passed to udf cannot be Unit. In your case it should be Double. Combining with the previous point a valid signature would be (Seq[Double], Seq[Double]) => Double.

colMod expects a single argument of type Column.

import org.apache.spark.sql.functions.array

colMod(array(preds.map(col): _*))

Your code is not NULL / null safe.

Collectives™ on Stack Overflow

scala/spark map returning an array instead of single value

1 Answer 1

Comments

Your Answer

Hot Network Questions

Collectives™ on Stack Overflow

1 Answer 1

Comments

Your Answer

Sign up or log in

Post as a guest

Related