stripe-archive / brushfire   0.7.2


Distributed decision tree ensemble learning in Scala

Scala versions: 2.12 2.11 2.10



Brushfire is a framework for distributed supervised learning of decision tree ensemble models in Scala.

The basic approach to distributed tree learning is inspired by Google's PLANET, but considerably generalized thanks to Scala's type parameterization and Algebird's aggregation abstractions.

Brushfire currently supports:

  • binary and multi-class classifiers
  • numeric features (discrete and continuous)
  • categorical features (including those with very high cardinality)
  • k-fold cross validation and random forests
  • chi-squared test as a measure of split quality
  • feature importance and brier scores
  • Scalding/Hadoop as a distributed computing platform

In the future we plan to add support for:

  • regression trees
  • CHAID-like multi-way splits
  • error-based pruning
  • many more ways to evaluate splits and trees
  • Spark and single-node in-memory platforms


Thanks for assistance and contributions:

Quick start

sbt brushfireScalding/assembly
cd example
cat iris.output/step_03

If it worked, you should see a JSON representation of 4 versions of a decision tree for classifying irises.

To use brushfire in your own SBT project, add the following to your build.sbt:

libraryDependencies += "com.stripe" %% "brushfire" % "0.6.3"

To use brushfire as a jar in your own Maven project, add the following to your POM file:


Using Brushfire with Scalding

The only distributed computing platform that Brushfire currently supports is Scalding, version 0.12 or later.

The simplest way to use Brushfire with Scalding is by subclassing TrainerJob and overriding trainer to return an instance of Trainer. Example:

import com.stripe.brushfire._
import com.stripe.brushfire.scalding._
import com.twitter.scalding._

class MyJob(args: Args) extends TrainerJob(args) {
  import JsonInjections._

  def trainer = ???

You should import either `JsonInjections` or `KryoInjections` to specify serialization in either JSON or base64-encoded Kryo, respectively; the former has the advantage of being human readable, the latter is more efficient, which can be important for very large trees.

To construct a `Trainer`, you need to pass it training data as a Scalding `TypedPipe` of Brushfire [Instance[K, V,T]]( objects. `Instance` looks like this:

case class Instance[K, V, T](id: String, timestamp: Long, features: Map[K, V], target: T)
  • The id should be unique for each instance.
  • If there's an associated observation time, it should be the timestamp. (Otherwise 0L is fine)
  • features is a Map from feature name (type K, usually String) to some value of type V. There's built-in implicit support for Int, Double, Boolean, and String types (with the assumption for Int and String that there is a small, finite number of possible values). If, as is common, you need to mix different feature types, see the section on Dispatched below.
  • the only built-in support for target currently is for Map[L,Long], where L represents some label type (for example Boolean for a binary classifier or String for multi-class). The Long values represent the weight for the instance, which is usually 1.


Instance("AS-2014", 1416168857L, Map("lat" -> 49.2, "long" -> 37.1, "altitude" -> 35000.0), Map(true -> 1L))

You also need to pass it a Sampler. Here are some samplers you might use:

One you have constructed a Trainer, you most likely want to call expandTimes(base: String, times: Int). This will build a new ensemble of trees from the training data and expand them times times, to depth times. At each step, the trees will be serialized to a directory (on HDFS, unless you're running in local mode) under base.

Fuller example:

import com.stripe.brushfire._
import com.stripe.brushfire.scalding._
import com.twitter.scalding._

class MyJob(args: Args) extends TrainerJob(args) {
  import JsonInjections._

  def trainingData: TypedPipe[Instance[K, V,T]] = ???
  def trainer = Trainer(trainingData, KFoldSampler(4)).expandTimes(args("output"), 5)

#In Memory Expansion

Having expanded as deep as you want using the distributed algorithm, you may wish to ask for further, in-memory expansion of any nodes that are sufficiently small at this point by calling expandSmallNodes(path: String, times: Int). By default, this will downsample every node to at most 10,000 instances of training data, and expand until they have fewer than 10 instances. You may need to tune this value, which you do by setting an implicit Stopper:

val implicit stopper = FrequencyStopper(10000, 10)
trainer.expandInMemory(args("output") + "/mem", 100)

Note that the distributed algorithm will *stop* expanding at the same instance count that the in-memory algorithm wants, ie, 10,000 instances by default.

# Dispatched

If you have mixed features, the recommended value type is `Dispatched[Int,String,Double,String]`, which requires your feature values to match any one of these four cases:

* `Ordinal(v: Int)` for numeric features with a reasonably small number of possible values
* `Nominal(v: String)` for categorical features with a reasonably small number of possible values
* `Continuous(v: Double)` for numeric features with a large or infinite number of possible values
* `Sparse(v: String)` for categorical features with a large or infinite number of possible values

Note that using `Sparse` and especially `Continuous` features will currently slow learning down considerably. (But on the other hand, if you try to use `Ordinal` or `Nominal` with a feature that has hundreds of thousands of unique values, it will be even slower, and then fail).

Example of a features map:

Map("age" -> Ordinal(35), "gender" -> Nominal("male"), "weight" -> Continuous(130.23), "name" -> Sparse("John"))

Extending Brushfire

Brushfire is designed to be extremely pluggable. Some ways you might want to extend it are (from simplest to most involved):

  • Adding a new sampling strategy, to get finer grained control over how instances are allocated to trees, or between the training set and the test set: define a new Sampler
  • Add a new evaluation strategy (such as log-likelihood or entropy): define a new Evaluator
  • Adding a new feature type, or a new way of binning an existing feature type (such as log-binning real numbers): define a new Splitter
  • Adding a new target type (such as real-valued targets for regression trees): define a new Evaluator, a new Stopper and quite likely also define a new Splitter for any continuous or sparse feature types you want to be able to use.
  • Add a new distributed computation platform: define a new equivalent of Trainer, idiomatically to the platform you're using. (There's no specific interface this should implement.)