ai.h2o.automl.targetencoding

Class TargetEncoder

java.lang.Object

ai.h2o.automl.targetencoding.TargetEncoder

public class TargetEncoder
extends java.lang.Object

Status: alpha version This is a core class for target encoding related logic. In general target encoding could be applied to three types of problems, namely: 1) Binary classification (supported) 2) Multi-class classification (not supported yet) 3) Regression (not supported yet) In order to differentiate between abovementioned types of problems at hand and enable algorithm to do encodings correctly user should explicitly set corresponding type for a response column: 1) Binary classification: response column should be of a categorical type with cardinality = 2 2) Multi-class: response column should be of a categorical type with cardinality > 2 3) Regression: response column should be of a numerical type Usage: see TargetEncodingTitanicBenchmark.java

Nested Class Summary

Nested Classes

Modifier and Type	Class and Description
static class	TargetEncoder.AddNoiseTask
static class	TargetEncoder.DataLeakageHandlingStrategy
static class	TargetEncoder.SubtractCurrentRowForLeaveOneOutTask

Constructor Summary

Constructors

Constructor and Description
TargetEncoder(java.lang.String[] columnNamesToEncode)
TargetEncoder(java.lang.String[] columnNamesToEncode, BlendingParams blendingParams)

Method Summary

Modifier and Type	Method and Description
water.fvec.Frame	applyTargetEncoding(water.fvec.Frame data, java.lang.String targetColumnName, java.util.Map<java.lang.String,water.fvec.Frame> targetEncodingMap, byte dataLeakageHandlingStrategy, boolean withBlendedAvg, boolean imputeNAs, long seed)
water.fvec.Frame	applyTargetEncoding(water.fvec.Frame data, java.lang.String targetColumnName, java.util.Map<java.lang.String,water.fvec.Frame> targetEncodingMap, byte dataLeakageHandlingStrategy, boolean withBlendedAvg, double noiseLevel, boolean imputeNAs, long seed)
water.fvec.Frame	applyTargetEncoding(water.fvec.Frame data, java.lang.String targetColumnName, java.util.Map<java.lang.String,water.fvec.Frame> targetEncodingMap, byte dataLeakageHandlingStrategy, java.lang.String foldColumn, boolean withBlendedAvg, boolean imputeNAs, long seed)
water.fvec.Frame	applyTargetEncoding(water.fvec.Frame data, java.lang.String targetColumnName, java.util.Map<java.lang.String,water.fvec.Frame> columnToEncodingMap, byte dataLeakageHandlingStrategy, java.lang.String foldColumnName, boolean withBlendedAvg, double noiseLevel, boolean imputeNAsWithNewCategory, long seed) Core method for applying pre-calculated encodings to the dataset.
java.util.Map<java.lang.String,water.fvec.Frame>	prepareEncodingMap(water.fvec.Frame data, java.lang.String targetColumnName, java.lang.String foldColumnName)
java.util.Map<java.lang.String,water.fvec.Frame>	prepareEncodingMap(water.fvec.Frame data, java.lang.String targetColumnName, java.lang.String foldColumnName, boolean imputeNAsWithNewCategory)

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

TargetEncoder

public TargetEncoder(java.lang.String[] columnNamesToEncode,
                     BlendingParams blendingParams)

Parameters:

columnNamesToEncode - names of columns to apply target encoding to

blendingParams -

TargetEncoder

public TargetEncoder(java.lang.String[] columnNamesToEncode)

Method Detail

prepareEncodingMap

public java.util.Map<java.lang.String,water.fvec.Frame> prepareEncodingMap(water.fvec.Frame data,
                                                                           java.lang.String targetColumnName,
                                                                           java.lang.String foldColumnName,
                                                                           boolean imputeNAsWithNewCategory)

Parameters:

targetColumnName - name of the target column

foldColumnName - name of the column that contains fold number the row is belong to

imputeNAsWithNewCategory - set to `true` to impute NAs with new category. // TODO probably we need to always set it to true bc we do not support null values on the right side of merge operation.

prepareEncodingMap

public java.util.Map<java.lang.String,water.fvec.Frame> prepareEncodingMap(water.fvec.Frame data,
                                                                           java.lang.String targetColumnName,
                                                                           java.lang.String foldColumnName)

applyTargetEncoding

public water.fvec.Frame applyTargetEncoding(water.fvec.Frame data,
                                            java.lang.String targetColumnName,
                                            java.util.Map<java.lang.String,water.fvec.Frame> columnToEncodingMap,
                                            byte dataLeakageHandlingStrategy,
                                            java.lang.String foldColumnName,
                                            boolean withBlendedAvg,
                                            double noiseLevel,
                                            boolean imputeNAsWithNewCategory,
                                            long seed)

Core method for applying pre-calculated encodings to the dataset. There are multiple overloaded methods that we will probably be able to get rid off if we are not going to expose Java API for TE. We can just stick to one signature that will suit internal representations of the AutoML's pipeline.

Parameters:

data - dataset that will be used as a base for creation of encodings .

targetColumnName - name of the column with respect to which we were computing encodings.

columnToEncodingMap - map of the prepared encodings with the keys being the names of the columns.

dataLeakageHandlingStrategy - see TargetEncoding.DataLeakageHandlingStrategy //TODO use common interface for stronger type safety.

foldColumnName - column's name that contains fold number the row is belong to.

withBlendedAvg - whether to apply blending or not.

noiseLevel - amount of noise to add to the final encodings.

imputeNAsWithNewCategory - set to `true` to impute NAs with new category.

seed - we might want to specify particular values for reproducibility in tests.

Returns:

copy of the `data` frame with encodings

applyTargetEncoding

public water.fvec.Frame applyTargetEncoding(water.fvec.Frame data,
                                            java.lang.String targetColumnName,
                                            java.util.Map<java.lang.String,water.fvec.Frame> targetEncodingMap,
                                            byte dataLeakageHandlingStrategy,
                                            java.lang.String foldColumn,
                                            boolean withBlendedAvg,
                                            boolean imputeNAs,
                                            long seed)

applyTargetEncoding

public water.fvec.Frame applyTargetEncoding(water.fvec.Frame data,
                                            java.lang.String targetColumnName,
                                            java.util.Map<java.lang.String,water.fvec.Frame> targetEncodingMap,
                                            byte dataLeakageHandlingStrategy,
                                            boolean withBlendedAvg,
                                            boolean imputeNAs,
                                            long seed)

applyTargetEncoding

public water.fvec.Frame applyTargetEncoding(water.fvec.Frame data,
                                            java.lang.String targetColumnName,
                                            java.util.Map<java.lang.String,water.fvec.Frame> targetEncodingMap,
                                            byte dataLeakageHandlingStrategy,
                                            boolean withBlendedAvg,
                                            double noiseLevel,
                                            boolean imputeNAs,
                                            long seed)