Machine Learning Library (MLlib) Programming Guide

MLlib is Spark’s scalable machine learning library consisting of common learning algorithms and utilities, including classification, regression, clustering, collaborative filtering, dimensionality reduction, as well as underlying optimization primitives, as outlined below:

MLlib is under active development. The APIs marked Experimental/DeveloperApi may change in future releases, and the migration guide below will explain all changes between releases.

spark.ml: The New ML Package

Spark 1.2 includes a new machine learning package called spark.ml, currently an alpha component but potentially a successor to spark.mllib. The spark.ml package aims to replace the old APIs with a cleaner, more uniform set of APIs which will help users create full machine learning pipelines.

See the spark.ml programming guide for more information on this package.

Users can use algorithms from either of the two packages, but APIs may differ. Currently, spark.ml offers a subset of the algorithms from spark.mllib.

Developers should contribute new algorithms to spark.mllib and can optionally contribute to spark.ml. See the spark.ml programming guide linked above for more details.

Dependencies

MLlib uses the linear algebra package Breeze, which depends on netlib-java, and jblas. netlib-java and jblas depend on native Fortran routines. You need to install the gfortran runtime library if it is not already present on your nodes. MLlib will throw a linking error if it cannot detect these libraries automatically. Due to license issues, we do not include netlib-java’s native libraries in MLlib’s dependency set under default settings. If no native library is available at runtime, you will see a warning message. To use native libraries from netlib-java, please build Spark with -Pnetlib-lgpl or include com.github.fommil.netlib:all:1.1.2 as a dependency of your project. If you want to use optimized BLAS/LAPACK libraries such as OpenBLAS, please link its shared libraries to /usr/lib/libblas.so.3 and /usr/lib/liblapack.so.3, respectively. BLAS/LAPACK libraries on worker nodes should be built without multithreading.

To use MLlib in Python, you will need NumPy version 1.4 or newer.


Migration Guide

From 1.1 to 1.2

The only API changes in MLlib v1.2 are in DecisionTree, which continues to be an experimental API in MLlib 1.2:

  1. (Breaking change) The Scala API for classification takes a named argument specifying the number of classes. In MLlib v1.1, this argument was called numClasses in Python and numClassesForClassification in Scala. In MLlib v1.2, the names are both set to numClasses. This numClasses parameter is specified either via Strategy or via DecisionTree static trainClassifier and trainRegressor methods.

  2. (Breaking change) The API for Node has changed. This should generally not affect user code, unless the user manually constructs decision trees (instead of using the trainClassifier or trainRegressor methods). The tree Node now includes more information, including the probability of the predicted label (for classification).

  3. Printing methods’ output has changed. The toString (Scala/Java) and __repr__ (Python) methods used to print the full model; they now print a summary. For the full model, use toDebugString.

Examples in the Spark distribution and examples in the Decision Trees Guide have been updated accordingly.

From 1.0 to 1.1

The only API changes in MLlib v1.1 are in DecisionTree, which continues to be an experimental API in MLlib 1.1:

  1. (Breaking change) The meaning of tree depth has been changed by 1 in order to match the implementations of trees in scikit-learn and in rpart. In MLlib v1.0, a depth-1 tree had 1 leaf node, and a depth-2 tree had 1 root node and 2 leaf nodes. In MLlib v1.1, a depth-0 tree has 1 leaf node, and a depth-1 tree has 1 root node and 2 leaf nodes. This depth is specified by the maxDepth parameter in Strategy or via DecisionTree static trainClassifier and trainRegressor methods.

  2. (Non-breaking change) We recommend using the newly added trainClassifier and trainRegressor methods to build a DecisionTree, rather than using the old parameter class Strategy. These new training methods explicitly separate classification and regression, and they replace specialized parameter types with simple String types.

Examples of the new, recommended trainClassifier and trainRegressor are given in the Decision Trees Guide.

From 0.9 to 1.0

In MLlib v1.0, we support both dense and sparse input in a unified way, which introduces a few breaking changes. If your data is sparse, please store it in a sparse format instead of dense to take advantage of sparsity in both storage and computation. Details are described below.

We used to represent a feature vector by Array[Double], which is replaced by Vector in v1.0. Algorithms that used to accept RDD[Array[Double]] now take RDD[Vector]. LabeledPoint is now a wrapper of (Double, Vector) instead of (Double, Array[Double]). Converting Array[Double] to Vector is straightforward:

import org.apache.spark.mllib.linalg.{Vector, Vectors}

val array: Array[Double] = ... // a double array
val vector: Vector = Vectors.dense(array) // a dense vector

Vectors provides factory methods to create sparse vectors.

Note: Scala imports scala.collection.immutable.Vector by default, so you have to import org.apache.spark.mllib.linalg.Vector explicitly to use MLlib’s Vector.

We used to represent a feature vector by double[], which is replaced by Vector in v1.0. Algorithms that used to accept RDD<double[]> now take RDD<Vector>. LabeledPoint is now a wrapper of (double, Vector) instead of (double, double[]). Converting double[] to Vector is straightforward:

import org.apache.spark.mllib.linalg.Vector;
import org.apache.spark.mllib.linalg.Vectors;

double[] array = ... // a double array
Vector vector = Vectors.dense(array); // a dense vector

Vectors provides factory methods to create sparse vectors.

We used to represent a labeled feature vector in a NumPy array, where the first entry corresponds to the label and the rest are features. This representation is replaced by class LabeledPoint, which takes both dense and sparse feature vectors.

from pyspark.mllib.linalg import SparseVector
from pyspark.mllib.regression import LabeledPoint

# Create a labeled point with a positive label and a dense feature vector.
pos = LabeledPoint(1.0, [1.0, 0.0, 3.0])

# Create a labeled point with a negative label and a sparse feature vector.
neg = LabeledPoint(0.0, SparseVector(3, [0, 2], [1.0, 3.0]))