Model evaluation

Without evaluation, a model is worth nothing as we don't know how accurately it performs. Therefore, we will now use the built-in BinaryClassificationEvaluator in order to assess prediction performance and a widely used measure called areaUnderROC (going into detail here is beyond the scope of this book):

import org.apache.spark.ml.evaluation.BinaryClassificationEvaluator
val evaluator = new BinaryClassificationEvaluator()

import org.apache.spark.ml.param.ParamMap
var evaluatorParamMap = ParamMap(evaluator.metricName -> "areaUnderROC")
var aucTraining = evaluator.evaluate(result, evaluatorParamMap)

As we can see, there is a built-in class called org.apache.spark.ml.evaluation.BinaryClassificationEvaluator and there are some other classes for other prediction use cases such as RegressionEvaluator or MuliclassClassificationEvaluator. The evaluator takes a parameter map--in this case, we are telling it to use the areaUnderROC metric--and finally, the evaluate method evaluates the result:

As we can see, areaUnderROC is 0.5424418446501833. An ideal classifier would return a score of one. So we are only doing a bit better than random guesses but, as already stated, the number of features that we are looking at is fairly limited.

In the previous example we are using the areaUnderROC metric which is used for evaluation of binary classifiers. There exist an abundance of other metrics used for different disciplines of machine learning such as accuracy, precision, recall and F1 score. The following provides a good overview http://www.cs.cornell.edu/courses/cs578/2003fa/performance_measures.pdf

This areaUnderROC is in fact a very bad value. Let's see if choosing better parameters for our RandomForest model increases this a bit in the next section.

Table of Contents for Model evaluation

Create new playlist

Sign In

Sign Up

Table of Contents for
Model evaluation