Obtain SHAP values from MOJO model

You can train the pipeline in Sparkling Water and get contributions from it or you can also get contributions from raw mojo. The following two sections describe how to achieve that.

Train model pipeline & get contributions

Obtaining SHAP values is possible only from H2OGBM, H2OXGBoost and H2ODRF pipeline wrappers and for regression or binomial problems.

To get SHAP values(=contributions) from H2OXGBoost model, please do:

Scala

First, let’s start Sparkling Shell as

./bin/sparkling-shell

Start H2O cluster inside the Spark environment

import org.apache.spark.h2o._
import java.net.URI
val hc = H2OContext.getOrCreate()

Parse the data using H2O and convert them to Spark Frame

val frame = new H2OFrame(new URI("https://raw.githubusercontent.com/h2oai/sparkling-water/master/examples/smalldata/prostate/prostate.csv"))
val sparkDF = hc.asSparkFrame(frame).withColumn("CAPSULE", $"CAPSULE" cast "string")
val Array(trainingDF, testingDF) = sparkDF.randomSplit(Array(0.8, 0.2))

Train the model. You can configure all the available XGBoost arguments using provided setters, such as the label column.

import ai.h2o.sparkling.ml.algos.H2OXGBoost
val estimator = new H2OXGBoost().setLabelCol("CAPSULE").setWithDetailedPredictionCol(true)
val model = estimator.fit(trainingDF)

The call setWithDetailedPredictionCol(true) tells Sparkling Water to create additional prediction column with additional prediction details, such as the contributions. The name of this column is by default “detailed_prediction” and can be modified via setDetailedPredictionCol setter.

Run Predictions

val predictions = model.transform(testingDF).show(false)

Show contributions

predictions.select("detailed_prediction.contribution").show()

Python

First, let’s start PySparkling Shell as

./bin/pysparkling

Start H2O cluster inside the Spark environment

from pysparkling import *
hc = H2OContext.getOrCreate()

Parse the data using H2O and convert them to Spark Frame

import h2o
frame = h2o.import_file("https://raw.githubusercontent.com/h2oai/sparkling-water/master/examples/smalldata/prostate/prostate.csv")
sparkDF = hc.asSparkFrame(frame)
sparkDF = sparkDF.withColumn("CAPSULE", sparkDF.CAPSULE.cast("string"))
[trainingDF, testingDF] = sparkDF.randomSplit([0.8, 0.2])

Train the model. You can configure all the available XGBoost arguments using provided setters or constructor parameters, such as the label column.

from pysparkling.ml import H2OXGBoost
estimator = H2OXGBoost(labelCol = "CAPSULE", withDetailedPredictionCol = True)
model = estimator.fit(trainingDF)

The parameter withDetailedPredictionCol = True tells Sparkling Water to create additional prediction column with additional prediction details, such as the contributions. The name of this column is by default “detailed_prediction” and can be modified via detailedPredictionCol parameter.

Run Predictions

model.transform(testingDF).show(truncate = False)

Show contributions

predictions.select("detailed_prediction.contributions").show()

Get Contributions from Raw MOJO

Obtaining SHAP values is possible only from MOJO’s generated for GBM, XGBoost and DRF and for regression or binomial problems. If you don’t need to train the model and just need to load existing mojo, there is no need to start H2OContext.

Scala

First, let’s start Sparkling Shell as

./bin/sparkling-shell

Parse the data using Spark

val testingDF = spark.read.option("header", "true").option("inferSchema", "true").csv("/path/to/testing/dataset.csv")

Load the existing MOJO and enable generation of contributions via the settings object.

import ai.h2o.sparkling.ml.models._

val path = "/path/to/mojo.zip"
val settings = H2OMOJOSettings(withDetailedPredictionCol = true)
val model = H2OMOJOModel.createFromMojo(path, settings)

Run Predictions

val predictions = model.transform(testingDF)

Show contributions

predictions.select("detailed_prediction.contributions").show()

Python

First, let’s start PySparkling Shell as

./bin/pysparkling

Parse the data using Spark

testingDF = spark.read.csv("/path/to/testing/dataset.csv", header=True, inferSchema=True)

Load the existing MOJO and enable generation of contributions via the settings object.

from pysparkling.ml import *

val path = '/path/to/mojo.zip'
settings = H2OMOJOSettings(withDetailedPredictionCol=True)
model = H2OMOJOModel.createFromMojo(path, settings)

Run Predictions

val predictions = model.transform(testingDF)

Show contributions

predictions.select("detailed_prediction.contributions").show()