This project has retired. For details please refer to its Attic page.

The classification template uses the Naive Bayes algorithm by default. You can easily add and use other MLlib classification algorithms. The following will demonstrate how to add the MLlib Random Forests algorithm into the engine.

You can find the complete modified source code here.

Create a new file RandomForestAlgorithm.scala

Locate src/main/scala/NaiveBayesAlgorithm.scala under your engine directory, which should be /MyClassification if you are following the Classification QuickStart. Copy NaiveBayesAlgorithm.scala and create a new file RandomForestAlgorithm.scala. You will modify this file and follow the instructions below to define a new RandomForestAlgorithm class.

Define the algorithm class and parameters

In 'RandomForestAlgorithm.scala', import the MLlib Random Forests algorithm by changing the following lines:

Original

1
2
import org.apache.spark.mllib.classification.NaiveBayes
import org.apache.spark.mllib.classification.NaiveBayesModel

Change to:

1
2
import org.apache.spark.mllib.tree.RandomForest // CHANGED
import org.apache.spark.mllib.tree.model.RandomForestModel // CHANGED

These are the necessary classes in order to use the MLLib's Random Forest algorithm.

Modify the AlgorithmParams class for the Random Forest algorithm:

1
2
3
4
5
6
7
8
9
// CHANGED
case class RandomForestAlgorithmParams(
  numClasses: Int,
  numTrees: Int,
  featureSubsetStrategy: String,
  impurity: String,
  maxDepth: Int,
  maxBins: Int
) extends Params

This class defines the parameters of the Random Forest algorithm (which later you can specify the value in engine.json). Please refer to MLlib documentation for the description and usage of these parameters.

Modify the NaiveBayesAlgorithm class to RandomForestAlgorithm. The changes are:

  • The new RandomForestAlgorithmParams class is used as parameter.
  • RandomForestModel is used in type parameter. This is the model returned by the Random Forest algorithm.
  • the train() function is modified and it returns the RandomForestModel instead of NaiveBayesModel.
  • the predict() function takes the RandomForestModel as input.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
// extends P2LAlgorithm because the MLlib's RandomForestModel doesn't
// contain RDD.
class RandomForestAlgorithm(val ap: RandomForestAlgorithmParams) // CHANGED
  extends P2LAlgorithm[PreparedData, RandomForestModel, // CHANGED
  Query, PredictedResult] {

  // CHANGED
  def train(sc: SparkContext, data: PreparedData): RandomForestModel = {
    // CHANGED
    // Empty categoricalFeaturesInfo indicates all features are continuous.
    val categoricalFeaturesInfo = Map[Int, Int]()
    RandomForest.trainClassifier(
      data.labeledPoints,
      ap.numClasses,
      categoricalFeaturesInfo,
      ap.numTrees,
      ap.featureSubsetStrategy,
      ap.impurity,
      ap.maxDepth,
      ap.maxBins)
  }

  def predict(
    model: RandomForestModel, // CHANGED
    query: Query): PredictedResult = {

    val label = model.predict(Vectors.dense(
      Array(query.attr0, query.attr1, query.attr2)
    ))
    PredictedResult(label)
  }

}

Note that the MLlib Random Forest algorithm takes the same training data as the Naive Bayes algorithm (ie, RDD[LabeledPoint]) so you don't need to modify the DataSource and PreparedData classes. If the new algorithm to be added requires different types of training data, then you need to modify these classes accordingly to accommodate your new algorithm.

Update Engine.scala

Modify the EngineFactory to add the new algorithm class RandomForestAlgorithm you just defined and give it a name "randomforest". The name will be used in engine.json to specify which algorithm to use.

1
2
3
4
5
6
7
8
9
10
object ClassificationEngine extends IEngineFactory {
  def apply() = {
    new Engine(
      classOf[DataSource],
      classOf[Preparator],
      Map("naive" -> classOf[NaiveBayesAlgorithm],
        "randomforest" -> classOf[RandomForestAlgorithm]), // ADDED
      classOf[Serving])
  }
}

This engine factory now returns an engine with two algorithms and they are named as "naive" and "randomforest" respectively.

Update engine.json

In order to use the new algorithm, you need to modify engine.json to specify the name of the algorithm and the parameters.

Update the engine.json to use randomforest:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
...
"algorithms": [
  {
    "name": "randomforest",
    "params": {
      "numClasses": 4,
      "numTrees": 5,
      "featureSubsetStrategy": "auto",
      "impurity": "gini",
      "maxDepth": 4,
      "maxBins": 100
    }
  }
]
...

The engine now uses MLlib Random Forests algorithm instead of the default Naive Bayes algorithm. You are ready to build, train and deploy the engine as described in quickstart.

1
2
3
$ pio build
$ pio train
$ pio deploy

To switch back using Naive Bayes algorithm, simply modify engine.json.