TransmogrifAI

TransmogrifAI

TransmogrifAI (pronounced trăns-mŏgˈrə-fī) is an AutoML library written in Scala that runs on top of Apache Spark. It was developed with a focus on accelerating machine learning developer productivity through machine learning automation, and an API that enforces compile-time type-safety, modularity, and reuse. Through automation, it achieves accuracies close to hand-tuned models with almost 100x reduction in time.

Use TransmogrifAI if you need a machine learning library to:

• Build production ready machine learning applications in hours, not months
• Build machine learning models without getting a Ph.D. in machine learning
• Build modular, reusable, strongly typed machine learning workflows

To understand the motivation behind TransmogrifAI check out these:

• Open Sourcing TransmogrifAI: Automated Machine Learning for Structured Data, a blog post by @snabar
• Meet TransmogrifAI, Open Source AutoML That Powers Einstein Predictions, a talk by @tovbinm
• Low Touch Machine Learning, a talk by @leahmcguire

Skip to Quick Start and Documentation.

Predicting Titanic Survivors with TransmogrifAI

The Titanic dataset is an often-cited dataset in the machine learning community. The goal is to build a machine learnt model that will predict survivors from the Titanic passenger manifest. Here is how you would build the model using TransmogrifAI:

import com.salesforce.op._
import com.salesforce.op.readers._
import com.salesforce.op.features._
import com.salesforce.op.features.types._
import com.salesforce.op.stages.impl.classification._
import org.apache.spark.SparkConf
import org.apache.spark.sql.SparkSession

implicit val spark = SparkSession.builder.config(new SparkConf()).getOrCreate()
import spark.implicits._

// Read Titanic data as a DataFrame
val passengersData = DataReaders.Simple.csvCase[Passenger](path = pathToData).readDataset().toDF()

// Extract response and predictor Features
val (survived, predictors) = FeatureBuilder.fromDataFrame[RealNN](passengersData, response = "survived")

// Automated feature engineering
val featureVector = predictors.transmogrify()

// Automated feature validation and selection
val checkedFeatures = survived.sanityCheck(featureVector, removeBadFeatures = true)

// Automated model selection
val pred = BinaryClassificationModelSelector().setInput(survived, checkedFeatures).getOutput()

// Setting up a TransmogrifAI workflow and training the model
val model = new OpWorkflow().setInputDataset(passengersData).setResultFeatures(pred).train()

println("Model summary:\n" + model.summaryPretty())

Model summary:

Evaluated Logistic Regression, Random Forest models with 3 folds and AuPR metric.
Evaluated 3 Logistic Regression models with AuPR between [0.6751930383321765, 0.7768725281794376]
Evaluated 16 Random Forest models with AuPR between [0.7781671467343991, 0.8104798040316159]

Selected model Random Forest classifier with parameters:
|-----------------------|--------------|
| Model Param           |     Value    |
|-----------------------|--------------|
| modelType             | RandomForest |
| featureSubsetStrategy |         auto |
| impurity              |         gini |
| maxBins               |           32 |
| maxDepth              |           12 |
| minInfoGain           |        0.001 |
| minInstancesPerNode   |           10 |
| numTrees              |           50 |
| subsamplingRate       |          1.0 |
|-----------------------|--------------|

Model evaluation metrics:
|-------------|--------------------|---------------------|
| Metric Name | Hold Out Set Value |  Training Set Value |
|-------------|--------------------|---------------------|
| Precision   |               0.85 |   0.773851590106007 |
| Recall      | 0.6538461538461539 |  0.6930379746835443 |
| F1          | 0.7391304347826088 |  0.7312186978297163 |
| AuROC       | 0.8821603927986905 |  0.8766642291593114 |
| AuPR        | 0.8225075757571668 |   0.850331080886535 |
| Error       | 0.1643835616438356 | 0.19682151589242053 |
| TP          |               17.0 |               219.0 |
| TN          |               44.0 |               438.0 |
| FP          |                3.0 |                64.0 |
| FN          |                9.0 |                97.0 |
|-------------|--------------------|---------------------|

Top model insights computed using correlation:
|-----------------------|----------------------|
| Top Positive Insights |      Correlation     |
|-----------------------|----------------------|
| sex = "female"        |   0.5177801026737666 |
| cabin = "OTHER"       |   0.3331391338844782 |
| pClass = 1            |   0.3059642953159715 |
|-----------------------|----------------------|
| Top Negative Insights |      Correlation     |
|-----------------------|----------------------|
| sex = "male"          |  -0.5100301587292186 |
| pClass = 3            |  -0.5075774968534326 |
| cabin = null          | -0.31463114463832633 |
|-----------------------|----------------------|

Top model insights computed using CramersV:
|-----------------------|----------------------|
|      Top Insights     |       CramersV       |
|-----------------------|----------------------|
| sex                   |    0.525557139885501 |
| embarked              |  0.31582347194683386 |
| age                   |  0.21582347194683386 |
|-----------------------|----------------------|

While this may seem a bit too magical, for those who want more control, TransmogrifAI also provides the flexibility to completely specify all the features being extracted and all the algorithms being applied in your ML pipeline. Visit our docs site for full documentation, getting started, examples, faq and other information.

Adding TransmogrifAI into your project

You can simply add TransmogrifAI as a regular dependency to an existing project. Start by picking TransmogrifAI version to match your project dependencies from the version matrix below (if not sure - take the stable version):

TransmogrifAI Version	Spark Version	Scala Version	Java Version
0.7.1 (unreleased, master), 0.7.0 (stable)	2.4	2.11	1.8
0.6.1, 0.6.0, 0.5.3, 0.5.2, 0.5.1, 0.5.0	2.3	2.11	1.8
0.4.0, 0.3.4	2.2	2.11	1.8

For Gradle in build.gradle add:

repositories {
    jcenter()
    mavenCentral()
}
dependencies {
    // TransmogrifAI core dependency
    compile 'com.salesforce.transmogrifai:transmogrifai-core_2.11:0.7.0'

    // TransmogrifAI pretrained models, e.g. OpenNLP POS/NER models etc. (optional)
    // compile 'com.salesforce.transmogrifai:transmogrifai-models_2.11:0.7.0'
}

For SBT in build.sbt add:

scalaVersion := "2.11.12"

resolvers += Resolver.jcenterRepo

// TransmogrifAI core dependency
libraryDependencies += "com.salesforce.transmogrifai" %% "transmogrifai-core" % "0.7.0"

// TransmogrifAI pretrained models, e.g. OpenNLP POS/NER models etc. (optional)
// libraryDependencies += "com.salesforce.transmogrifai" %% "transmogrifai-models" % "0.7.0"

Then import TransmogrifAI into your code:

// TransmogrifAI functionality: feature types, feature builders, feature dsl, readers, aggregators etc.
import com.salesforce.op._
import com.salesforce.op.aggregators._
import com.salesforce.op.features._
import com.salesforce.op.features.types._
import com.salesforce.op.readers._

// Spark enrichments (optional)
import com.salesforce.op.utils.spark.RichDataset._
import com.salesforce.op.utils.spark.RichRDD._
import com.salesforce.op.utils.spark.RichRow._
import com.salesforce.op.utils.spark.RichMetadata._
import com.salesforce.op.utils.spark.RichStructType._

Quick Start and Documentation

Visit our docs site for full documentation, getting started, examples, faq and other information.

See scaladoc for the programming API.

Authors

• Kevin Moore @jauntbox
• Kin Fai Kan @kinfaikan
• Leah McGuire @leahmcguire
• Matthew Tovbin @tovbinm
• Max Ovsiankin @maxov
• Michael Loh @mikeloh77
• Michael Weil @michaelweilsalesforce
• Shubha Nabar @snabar
• Vitaly Gordon @vitalyg
• Vlad Patryshev @vpatryshev

Internal Contributors (prior to release)

• Chris Rupley @crupley
• Chris Wu @cjwooo
• Eric Wayman @ericwayman
• Felipe Oliveira @feliperazeek
• Gera Shegalov @gerashegalov
• Jean-Marc Soumet @ajmssc
• Marco Vivero @marcovivero
• Mario Rodriguez @mrodriguezsfiq
• Mayukh Bhaowal @mayukhb
• Minh-An Quinn @minhanquinn
• Nicolas Drizard @nicodri
• Oleg Gusak @ogusak
• Patrick Framption @tricktrap
• Ryle Goehausen @ryleg
• Sanmitra Ijeri @sanmitra
• Sky Chen @almandsky
• Sophie Xiaodan Sun @sxd929
• Till Bergmann @tillbe
• Xiaoqian Liu @wingsrc

License

BSD 3-Clause © Salesforce.com, Inc.