Machine Learning Library (MLlib)

MLlib is a Spark implementation of some common machine learning (ML) functionality, as well associated tests and data generators. MLlib currently supports four common types of machine learning problem settings, namely, binary classification, regression, clustering and collaborative filtering, as well as an underlying gradient descent optimization primitive. This guide will outline the functionality supported in MLlib and also provides an example of invoking MLlib.

Dependencies

MLlib uses the jblas linear algebra library, which itself depends on native Fortran routines. You may 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.

Binary Classification

Binary classification is a supervised learning problem in which we want to classify entities into one of two distinct categories or labels, e.g., predicting whether or not emails are spam. This problem involves executing a learning Algorithm on a set of labeled examples, i.e., a set of entities represented via (numerical) features along with underlying category labels. The algorithm returns a trained Model that can predict the label for new entities for which the underlying label is unknown.

MLlib currently supports two standard model families for binary classification, namely Linear Support Vector Machines (SVMs) and Logistic Regression, along with L1 and L2 regularized variants of each model family. The training algorithms all leverage an underlying gradient descent primitive (described below), and take as input a regularization parameter (regParam) along with various parameters associated with gradient descent (stepSize, numIterations, miniBatchFraction).

The following code snippet illustrates how to load a sample dataset, execute a training algorithm on this training data using a static method in the algorithm object, and make predictions with the resulting model to compute the training error.

import org.apache.spark.SparkContext
import org.apache.spark.mllib.classification.SVMWithSGD
import org.apache.spark.mllib.regression.LabeledPoint

// Load and parse the data file
val data = sc.textFile("mllib/data/sample_svm_data.txt")
val parsedData = data.map { line =>
  val parts = line.split(' ')
  LabeledPoint(parts(0).toDouble, parts.tail.map(x => x.toDouble).toArray)
}

// Run training algorithm
val numIterations = 20
val model = SVMWithSGD.train(parsedData, numIterations)
 
// Evaluate model on training examples and compute training error
val labelAndPreds = parsedData.map { point =>
  val prediction = model.predict(point.features)
  (point.label, prediction)
}
val trainErr = labelAndPreds.filter(r => r._1 != r._2).count.toDouble / parsedData.count
println("trainError = " + trainErr)

The SVMWithSGD.train() method by default performs L2 regularization with the regularization parameter set to 1.0. If we want to configure this algorithm, we can customize SVMWithSGD further by creating a new object directly and calling setter methods. All other MLlib algorithms support customization in this way as well. For example, the following code produces an L1 regularized variant of SVMs with regularization parameter set to 0.1, and runs the training algorithm for 200 iterations. 

import org.apache.spark.mllib.optimization.L1Updater

val svmAlg = new SVMWithSGD()
svmAlg.optimizer.setNumIterations(200)
  .setRegParam(0.1)
  .setUpdater(new L1Updater)
val modelL1 = svmAlg.run(parsedData)

Both of the code snippets above can be executed in spark-shell to generate a classifier for the provided dataset.

Available algorithms for binary classification:

Linear Regression

Linear regression is another classical supervised learning setting. In this problem, each entity is associated with a real-valued label (as opposed to a binary label as in binary classification), and we want to predict labels as closely as possible given numerical features representing entities. MLlib supports linear regression as well as L1 (lasso) and L2 (ridge) regularized variants. The regression algorithms in MLlib also leverage the underlying gradient descent primitive (described below), and have the same parameters as the binary classification algorithms described above.

Available algorithms for linear regression:

Clustering

Clustering is an unsupervised learning problem whereby we aim to group subsets of entities with one another based on some notion of similarity. Clustering is often used for exploratory analysis and/or as a component of a hierarchical supervised learning pipeline (in which distinct classifiers or regression models are trained for each cluster). MLlib supports k-means clustering, arguably the most commonly used clustering approach that clusters the data points into k clusters. The MLlib implementation includes a parallelized variant of the k-means++ method called kmeans||. The implementation in MLlib has the following parameters:

Available algorithms for clustering:

Collaborative Filtering

Collaborative filtering is commonly used for recommender systems. These techniques aim to fill in the missing entries of a user-product association matrix. MLlib currently supports model-based collaborative filtering, in which users and products are described by a small set of latent factors that can be used to predict missing entries. In particular, we implement the alternating least squares (ALS) algorithm to learn these latent factors. The implementation in MLlib has the following parameters:

Available algorithms for collaborative filtering:

Gradient Descent Primitive

Gradient descent (along with stochastic variants thereof) are first-order optimization methods that are well-suited for large-scale and distributed computation. Gradient descent methods aim to find a local minimum of a function by iteratively taking steps in the direction of the negative gradient of the function at the current point, i.e., the current parameter value. Gradient descent is included as a low-level primitive in MLlib, upon which various ML algorithms are developed, and has the following parameters:

Available algorithms for gradient descent: