spark-instrumented-optimizer/docs/ml-guide.md

layout	title	displayTitle
global	Overview: estimators, transformers and pipelines - spark.ml	Overview: estimators, transformers and pipelines - spark.ml

\[ \newcommand{\R}{\mathbb{R}} \newcommand{\E}{\mathbb{E}} \newcommand{\x}{\mathbf{x}} \newcommand{\y}{\mathbf{y}} \newcommand{\wv}{\mathbf{w}} \newcommand{\av}{\mathbf{\alpha}} \newcommand{\bv}{\mathbf{b}} \newcommand{\N}{\mathbb{N}} \newcommand{\id}{\mathbf{I}} \newcommand{\ind}{\mathbf{1}} \newcommand{\0}{\mathbf{0}} \newcommand{\unit}{\mathbf{e}} \newcommand{\one}{\mathbf{1}} \newcommand{\zero}{\mathbf{0}} \]

The spark.ml package aims to provide a uniform set of high-level APIs built on top of DataFrames that help users create and tune practical machine learning pipelines. See the algorithm guides section below for guides on sub-packages of spark.ml, including feature transformers unique to the Pipelines API, ensembles, and more.

Table of contents

• This will become a table of contents (this text will be scraped). {:toc}

Main concepts in Pipelines

Spark ML standardizes APIs for machine learning algorithms to make it easier to combine multiple algorithms into a single pipeline, or workflow. This section covers the key concepts introduced by the Spark ML API, where the pipeline concept is mostly inspired by the scikit-learn project.

• DataFrame: Spark ML uses DataFrame from Spark SQL as an ML dataset, which can hold a variety of data types. E.g., a DataFrame could have different columns storing text, feature vectors, true labels, and predictions.

• Transformer: A Transformer is an algorithm which can transform one DataFrame into another DataFrame. E.g., an ML model is a Transformer which transforms DataFrame with features into a DataFrame with predictions.

• Estimator: An Estimator is an algorithm which can be fit on a DataFrame to produce a Transformer. E.g., a learning algorithm is an Estimator which trains on a DataFrame and produces a model.

• Pipeline: A Pipeline chains multiple Transformers and Estimators together to specify an ML workflow.

• Parameter: All Transformers and Estimators now share a common API for specifying parameters.

DataFrame

Machine learning can be applied to a wide variety of data types, such as vectors, text, images, and structured data. Spark ML adopts the DataFrame from Spark SQL in order to support a variety of data types.

DataFrame supports many basic and structured types; see the Spark SQL datatype reference for a list of supported types. In addition to the types listed in the Spark SQL guide, DataFrame can use ML Vector types.

A DataFrame can be created either implicitly or explicitly from a regular RDD. See the code examples below and the Spark SQL programming guide for examples.

Columns in a DataFrame are named. The code examples below use names such as "text," "features," and "label."

Pipeline components

Transformers

A Transformer is an abstraction that includes feature transformers and learned models. Technically, a Transformer implements a method transform(), which converts one DataFrame into another, generally by appending one or more columns. For example:

• A feature transformer might take a DataFrame, read a column (e.g., text), map it into a new column (e.g., feature vectors), and output a new DataFrame with the mapped column appended.
• A learning model might take a DataFrame, read the column containing feature vectors, predict the label for each feature vector, and output a new DataFrame with predicted labels appended as a column.

Estimators

An Estimator abstracts the concept of a learning algorithm or any algorithm that fits or trains on data. Technically, an Estimator implements a method fit(), which accepts a DataFrame and produces a Model, which is a Transformer. For example, a learning algorithm such as LogisticRegression is an Estimator, and calling fit() trains a LogisticRegressionModel, which is a Model and hence a Transformer.

Properties of pipeline components

Transformer.transform()s and Estimator.fit()s are both stateless. In the future, stateful algorithms may be supported via alternative concepts.

Each instance of a Transformer or Estimator has a unique ID, which is useful in specifying parameters (discussed below).

Pipeline

In machine learning, it is common to run a sequence of algorithms to process and learn from data. E.g., a simple text document processing workflow might include several stages:

• Split each document's text into words.
• Convert each document's words into a numerical feature vector.
• Learn a prediction model using the feature vectors and labels.

Spark ML represents such a workflow as a Pipeline, which consists of a sequence of PipelineStages (Transformers and Estimators) to be run in a specific order. We will use this simple workflow as a running example in this section.

How it works

A Pipeline is specified as a sequence of stages, and each stage is either a Transformer or an Estimator. These stages are run in order, and the input DataFrame is transformed as it passes through each stage. For Transformer stages, the transform() method is called on the DataFrame. For Estimator stages, the fit() method is called to produce a Transformer (which becomes part of the PipelineModel, or fitted Pipeline), and that Transformer's transform() method is called on the DataFrame.

We illustrate this for the simple text document workflow. The figure below is for the training time usage of a Pipeline.

Above, the top row represents a Pipeline with three stages. The first two (Tokenizer and HashingTF) are Transformers (blue), and the third (LogisticRegression) is an Estimator (red). The bottom row represents data flowing through the pipeline, where cylinders indicate DataFrames. The Pipeline.fit() method is called on the original DataFrame, which has raw text documents and labels. The Tokenizer.transform() method splits the raw text documents into words, adding a new column with words to the DataFrame. The HashingTF.transform() method converts the words column into feature vectors, adding a new column with those vectors to the DataFrame. Now, since LogisticRegression is an Estimator, the Pipeline first calls LogisticRegression.fit() to produce a LogisticRegressionModel. If the Pipeline had more stages, it would call the LogisticRegressionModel's transform() method on the DataFrame before passing the DataFrame to the next stage.

A Pipeline is an Estimator. Thus, after a Pipeline's fit() method runs, it produces a PipelineModel, which is a Transformer. This PipelineModel is used at test time; the figure below illustrates this usage.

In the figure above, the PipelineModel has the same number of stages as the original Pipeline, but all Estimators in the original Pipeline have become Transformers. When the PipelineModel's transform() method is called on a test dataset, the data are passed through the fitted pipeline in order. Each stage's transform() method updates the dataset and passes it to the next stage.

Pipelines and PipelineModels help to ensure that training and test data go through identical feature processing steps.

Details

DAG Pipelines: A Pipeline's stages are specified as an ordered array. The examples given here are all for linear Pipelines, i.e., Pipelines in which each stage uses data produced by the previous stage. It is possible to create non-linear Pipelines as long as the data flow graph forms a Directed Acyclic Graph (DAG). This graph is currently specified implicitly based on the input and output column names of each stage (generally specified as parameters). If the Pipeline forms a DAG, then the stages must be specified in topological order.

Runtime checking: Since Pipelines can operate on DataFrames with varied types, they cannot use compile-time type checking. Pipelines and PipelineModels instead do runtime checking before actually running the Pipeline. This type checking is done using the DataFrame schema, a description of the data types of columns in the DataFrame.

Unique Pipeline stages: A Pipeline's stages should be unique instances. E.g., the same instance myHashingTF should not be inserted into the Pipeline twice since Pipeline stages must have unique IDs. However, different instances myHashingTF1 and myHashingTF2 (both of type HashingTF) can be put into the same Pipeline since different instances will be created with different IDs.

Parameters

Spark ML Estimators and Transformers use a uniform API for specifying parameters.

A Param is a named parameter with self-contained documentation. A ParamMap is a set of (parameter, value) pairs.

There are two main ways to pass parameters to an algorithm:

1. Set parameters for an instance. E.g., if lr is an instance of LogisticRegression, one could call lr.setMaxIter(10) to make lr.fit() use at most 10 iterations. This API resembles the API used in spark.mllib package.
2. Pass a ParamMap to fit() or transform(). Any parameters in the ParamMap will override parameters previously specified via setter methods.

Parameters belong to specific instances of Estimators and Transformers. For example, if we have two LogisticRegression instances lr1 and lr2, then we can build a ParamMap with both maxIter parameters specified: ParamMap(lr1.maxIter -> 10, lr2.maxIter -> 20). This is useful if there are two algorithms with the maxIter parameter in a Pipeline.

Saving and Loading Pipelines

Often times it is worth it to save a model or a pipeline to disk for later use. In Spark 1.6, a model import/export functionality was added to the Pipeline API. Most basic transformers are supported as well as some of the more basic ML models. Please refer to the algorithm's API documentation to see if saving and loading is supported.

Code examples

This section gives code examples illustrating the functionality discussed above. For more info, please refer to the API documentation (Scala, Java, and Python). Some Spark ML algorithms are wrappers for spark.mllib algorithms, and the MLlib programming guide has details on specific algorithms.

Example: Estimator, Transformer, and Param

This example covers the concepts of Estimator, Transformer, and Param.

{% include_example scala/org/apache/spark/examples/ml/EstimatorTransformerParamExample.scala %}

{% include_example java/org/apache/spark/examples/ml/JavaEstimatorTransformerParamExample.java %}

{% include_example python/ml/estimator_transformer_param_example.py %}

Example: Pipeline

This example follows the simple text document Pipeline illustrated in the figures above.

{% include_example scala/org/apache/spark/examples/ml/PipelineExample.scala %}

{% include_example java/org/apache/spark/examples/ml/JavaPipelineExample.java %}

{% include_example python/ml/pipeline_example.py %}

Example: model selection via cross-validation

An important task in ML is model selection, or using data to find the best model or parameters for a given task. This is also called tuning. Pipelines facilitate model selection by making it easy to tune an entire Pipeline at once, rather than tuning each element in the Pipeline separately.

Currently, spark.ml supports model selection using the CrossValidator class, which takes an Estimator, a set of ParamMaps, and an Evaluator. CrossValidator begins by splitting the dataset into a set of folds which are used as separate training and test datasets; e.g., with $k=3$ folds, CrossValidator will generate 3 (training, test) dataset pairs, each of which uses 2/3 of the data for training and 1/3 for testing. CrossValidator iterates through the set of ParamMaps. For each ParamMap, it trains the given Estimator and evaluates it using the given Evaluator.

The Evaluator can be a RegressionEvaluator for regression problems, a BinaryClassificationEvaluator for binary data, or a MultiClassClassificationEvaluator for multiclass problems. The default metric used to choose the best ParamMap can be overridden by the setMetricName method in each of these evaluators.

The ParamMap which produces the best evaluation metric (averaged over the $k$ folds) is selected as the best model. CrossValidator finally fits the Estimator using the best ParamMap and the entire dataset.

The following example demonstrates using CrossValidator to select from a grid of parameters. To help construct the parameter grid, we use the ParamGridBuilder utility.

Note that cross-validation over a grid of parameters is expensive. E.g., in the example below, the parameter grid has 3 values for hashingTF.numFeatures and 2 values for lr.regParam, and CrossValidator uses 2 folds. This multiplies out to $(3 \times 2) \times 2 = 12$ different models being trained. In realistic settings, it can be common to try many more parameters and use more folds ($k=3$ and $k=10$ are common). In other words, using CrossValidator can be very expensive. However, it is also a well-established method for choosing parameters which is more statistically sound than heuristic hand-tuning.

{% include_example scala/org/apache/spark/examples/ml/ModelSelectionViaCrossValidationExample.scala %}

{% include_example java/org/apache/spark/examples/ml/JavaModelSelectionViaCrossValidationExample.java %}

{% include_example python/ml/cross_validator.py %}

Example: model selection via train validation split

In addition to CrossValidator Spark also offers TrainValidationSplit for hyper-parameter tuning. TrainValidationSplit only evaluates each combination of parameters once as opposed to k times in case of CrossValidator. It is therefore less expensive, but will not produce as reliable results when the training dataset is not sufficiently large.

TrainValidationSplit takes an Estimator, a set of ParamMaps provided in the estimatorParamMaps parameter, and an Evaluator. It begins by splitting the dataset into two parts using trainRatio parameter which are used as separate training and test datasets. For example with $trainRatio=0.75$ (default), TrainValidationSplit will generate a training and test dataset pair where 75% of the data is used for training and 25% for validation. Similar to CrossValidator, TrainValidationSplit also iterates through the set of ParamMaps. For each combination of parameters, it trains the given Estimator and evaluates it using the given Evaluator. The ParamMap which produces the best evaluation metric is selected as the best option. TrainValidationSplit finally fits the Estimator using the best ParamMap and the entire dataset.

{% include_example scala/org/apache/spark/examples/ml/ModelSelectionViaTrainValidationSplitExample.scala %}

{% include_example java/org/apache/spark/examples/ml/JavaModelSelectionViaTrainValidationSplitExample.java %}

{% include_example python/ml/train_validation_split.py %}