# AutoQuant is a package for quickly creating high quality visualizations under a common and easy api.
# Copyright (C) <year> <name of author>
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero General Public License as
# published by the Free Software Foundation, either version 3 of the
# License, or (at your option) any later version.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Affero General Public License for more details.
#
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
#' @title AutoH2oGLMRegression
#'
#' @description AutoH2oGLMis an automated H2O modeling framework with grid-tuning and model evaluation that runs a variety of steps. First, the function will run a random grid tune over N number of models and find which model is the best (a default model is always included in that set). Once the model is identified and built, several other outputs are generated: validation data with predictions, evaluation plot, evaluation boxplot, evaluation metrics, variable importance, partial dependence calibration plots, partial dependence calibration box plots, and column names used in model fitting.
#'
#' @author Adrian Antico
#' @family Automated Supervised Learning - Regression
#'
#' @param OutputSelection You can select what type of output you want returned. Choose from "EvalMetrics", "Score_TrainData", "h2o.explain"
#' @param data This is your data set for training and testing your model
#' @param TrainOnFull Set to TRUE to train on full data
#' @param ValidationData This is your holdout data set used in modeling either refine your hyperparameters.
#' @param TestData This is your holdout data set. Catboost using both training and validation data in the training process so you should evaluate out of sample performance with this data set.
#' @param TargetColumnName Either supply the target column name OR the column number where the target is located (but not mixed types).
#' @param FeatureColNames Either supply the feature column names OR the column number where the target is located (but not mixed types)
#' @param RandomColNumbers Random effects column number indicies. You can also pass character names of the columns.
#' @param InteractionColNumbers Column numbers of the features you want to be pairwise interacted
#' @param WeightsColumn Column name of a weights column
#' @param TransformNumericColumns Set to NULL to do nothing; otherwise supply the column names of numeric variables you want transformed
#' @param Methods Choose from "YeoJohnson", "BoxCox", "Asinh", "Log", "LogPlus1", "Sqrt", "Asin", or "Logit". If more than one is selected, the one with the best normalization pearson statistic will be used. Identity is automatically selected and compared.
#' @param eval_metric This is the metric used to identify best grid tuned model. Choose from "MSE", "RMSE", "MAE", "RMSLE"
#' @param GridTune Set to TRUE to run a grid tuning procedure. Set a number in MaxModelsInGrid to tell the procedure how many models you want to test.
#' @param GridStrategy "RandomDiscrete" or "Cartesian"
#' @param MaxRunTimeSecs Max run time in seconds
#' @param StoppingRounds Iterations in grid tuning
#' @param MaxModelsInGrid Number of models to test from grid options (1080 total possible options)
#' @param MaxMem Set the maximum amount of memory you'd like to dedicate to the model run. E.g. "32G"
#' @param NThreads Set the number of threads you want to dedicate to the model building
#' @param model_path A character string of your path file to where you want your output saved
#' @param metadata_path A character string of your path file to where you want your model evaluation output saved. If left NULL, all output will be saved to model_path.
#' @param ModelID A character string to name your model and output
#' @param NumOfParDepPlots Tell the function the number of partial dependence calibration plots you want to create. Calibration boxplots will only be created for numerical features (not dummy variables)
#' @param ReturnModelObjects Set to TRUE to output all modeling objects (E.g. plots and evaluation metrics)
#' @param SaveModelObjects Set to TRUE to return all modeling objects to your environment
#' @param SaveInfoToPDF Set to TRUE to save insights to PDF
#' @param IfSaveModel Set to "mojo" to save a mojo file, otherwise "standard" to save a regular H2O model object
#' @param H2OStartUp Defaults to TRUE which means H2O will be started inside the function
#' @param H2OShutdown Set to TRUE to shutdown H2O inside the function
#' @param DebugMode Set to TRUE to print out steps to screen
#' @param Distribution "AUTO", "gaussian", "poisson", "gamma", "tweedie", "negativebinomial"
#' @param Link "family_default", "identity", "log", "inverse", "tweedie"
#' @param TweedieLinkPower 1, 2, 3 for Poisson, Gamma, and Gaussian
#' @param TweedieVariancePower See h2o docs for background
#' @param RandomDistribution Random effects family. Defaults NULL, otherwise it will run a hierarchical glm
#' @param RandomLink Random effects link. Defaults NULL, otherwise it will run a hierarchical glm
#' @param Solver Default "AUTO". Options include "IRLSM", "L_BFGS", "COORDINATE_DESCENT_NAIVE", "COORDINATE_DESCENT", "GRADIENT_DESCENT_LH", "GRADIENT_DESCENT_SQERR"
#' @param Alpha Default 0.5 Otherwise supply a value between 0 and 1. 1 is equivalent to Lasso regression. 0 is equivalent to Ridge regression. Inbetween for a blend of the two.
#' @param Lambda Default NULL. Regularization strength.
#' @param LambdaSearch Default FALSE.
#' @param NLambdas Default -1
#' @param Standardize Default TRUE. Standardize numerical columns
#' @param RemoveCollinearColumns Default FALSE. Removes some of the linearly dependent columns
#' @param InterceptInclude Default TRUE
#' @param NonNegativeCoefficients Default FALSE
#' @examples
#' \donttest{
#' # Create some dummy correlated data
#' data <- AutoQuant::FakeDataGenerator(
#' Correlation = 0.85,
#' N = 1000,
#' ID = 2,
#' ZIP = 0,
#' AddDate = FALSE,
#' Classification = FALSE,
#' MultiClass = FALSE)
#'
#' # Run function
#' TestModel <- AutoQuant::AutoH2oGLMRegression(
#'
#' # Compute management
#' MaxMem = {gc();paste0(as.character(floor(as.numeric(system("awk '/MemFree/ {print $2}' /proc/meminfo", intern=TRUE)) / 1000000)),"G")},
#' NThreads = max(1, parallel::detectCores()-2),
#' H2OShutdown = TRUE,
#' H2OStartUp = TRUE,
#' IfSaveModel = "mojo",
#'
#' # Model evaluation:
#' eval_metric = "RMSE",
#' NumOfParDepPlots = 3,
#'
#' # Metadata arguments
#' OutputSelection = c("EvalMetrics", "Score_TrainData"),
#' model_path = NULL,
#' metadata_path = NULL,
#' ModelID = "FirstModel",
#' ReturnModelObjects = TRUE,
#' SaveModelObjects = FALSE,
#' SaveInfoToPDF = FALSE,
#' DebugMode = FALSE,
#'
#' # Data arguments:
#' data = data,
#' TrainOnFull = FALSE,
#' ValidationData = NULL,
#' TestData = NULL,
#' TargetColumnName = "Adrian",
#' FeatureColNames = names(data)[!names(data) %in%
#' c("IDcol_1", "IDcol_2","Adrian")],
#' RandomColNumbers = NULL,
#' InteractionColNumbers = NULL,
#' WeightsColumn = NULL,
#' TransformNumericColumns = NULL,
#' Methods = c("Asinh", "Asin", "Log", "LogPlus1", "Sqrt", "Logit"),
#'
#' # Model args
#' GridTune = FALSE,
#' GridStrategy = "Cartesian",
#' StoppingRounds = 10,
#' MaxRunTimeSecs = 3600 * 24 * 7,
#' MaxModelsInGrid = 10,
#' Distribution = "gaussian",
#' Link = "identity",
#' TweedieLinkPower = NULL,
#' TweedieVariancePower = NULL,
#' RandomDistribution = NULL,
#' RandomLink = NULL,
#' Solver = "AUTO",
#' Alpha = 0.5,
#' Lambda = NULL,
#' LambdaSearch = FALSE,
#' NLambdas = -1,
#' Standardize = TRUE,
#' RemoveCollinearColumns = FALSE,
#' InterceptInclude = TRUE,
#' NonNegativeCoefficients = FALSE)
#' }
#' @return Saves to file and returned in list: VariableImportance.csv, Model, ValidationData.csv, EvalutionPlot.png, EvalutionBoxPlot.png, EvaluationMetrics.csv, ParDepPlots.R a named list of features with partial dependence calibration plots, ParDepBoxPlots.R, GridCollect, GridList, and Transformation metadata
#' @export
AutoH2oGLMRegression <- function(OutputSelection = c("EvalMetrics", "Score_TrainData"),
data = NULL,
TrainOnFull = FALSE,
ValidationData = NULL,
TestData = NULL,
TargetColumnName = NULL,
FeatureColNames = NULL,
RandomColNumbers = NULL,
InteractionColNumbers = NULL,
WeightsColumn = NULL,
MaxMem = {gc();paste0(as.character(floor(as.numeric(system("awk '/MemFree/ {print $2}' /proc/meminfo", intern=TRUE)) / 1000000)),"G")},
NThreads = max(1, parallel::detectCores()-2),
ModelID = "FirstModel",
ReturnModelObjects = TRUE,
model_path = NULL,
metadata_path = NULL,
SaveModelObjects = FALSE,
SaveInfoToPDF = FALSE,
IfSaveModel = "mojo",
H2OShutdown = TRUE,
H2OStartUp = TRUE,
DebugMode = FALSE,
TransformNumericColumns = NULL,
Methods = c("BoxCox", "Asinh", "Log", "LogPlus1", "Sqrt", "Asin", "Logit"),
NumOfParDepPlots = 3,
GridTune = FALSE,
GridStrategy = "Cartesian",
StoppingRounds = 10,
MaxRunTimeSecs = 3600 * 24 * 7,
MaxModelsInGrid = 2,
Distribution = "gaussian",
Link = "identity",
TweedieLinkPower = NULL,
TweedieVariancePower = NULL,
eval_metric = "RMSE",
RandomDistribution = NULL,
RandomLink = NULL,
Solver = "AUTO",
Alpha = 0.5,
Lambda = NULL,
LambdaSearch = FALSE,
NLambdas = -1,
Standardize = TRUE,
RemoveCollinearColumns = FALSE,
InterceptInclude = TRUE,
NonNegativeCoefficients = FALSE) {
# Check Arguments ----
if(!(tolower(eval_metric) %chin% c("mse", "rmse", "mae", "rmsle"))) stop("eval_metric not in MSE, RMSE, MAE, RMSLE")
if(!GridTune %in% c(TRUE, FALSE)) stop("GridTune needs to be TRUE or FALSE")
if(MaxModelsInGrid < 1 && GridTune) stop("MaxModelsInGrid needs to be at least 1")
if(!is.null(model_path)) if(!is.character(model_path)) stop("model_path needs to be a character type")
if(!is.null(metadata_path)) if(!is.character(metadata_path)) stop("metadata_path needs to be a character type")
if(!is.character(ModelID) && !is.null(ModelID)) stop("ModelID needs to be a character type")
if(NumOfParDepPlots < 0) stop("NumOfParDepPlots needs to be a positive number")
if(!(ReturnModelObjects %in% c(TRUE, FALSE))) stop("ReturnModelObjects needs to be TRUE or FALSE")
if(!(SaveModelObjects %in% c(TRUE, FALSE))) stop("SaveModelObjects needs to be TRUE or FALSE")
if(length(RandomColNumbers) > 0L && !is.numeric(RandomColNumbers)) {
RandomColNumbers <- which(names(data) %in% RandomColNumbers)
if(length(ValidationData) > 0L) {
TrainData <- data.table::rbindlist(list(TrainData, ValidationData), use.names = TRUE, fill = TRUE)
ValidationData <- NULL
}
}
# Grab all official parameters and their evaluated arguments
ArgsList <- c(as.list(environment()))
ArgsList[['data']] <- NULL
ArgsList[['ValidationData']] <- NULL
ArgsList[['TestData']] <- NULL
ArgsList[['Algo']] <- "H2OGLM"
ArgsList[['TargetType']] <- "Regression"
ArgsList[['PredictionColumnName']] <- "Predict"
if(SaveModelObjects) {
if(!is.null(metadata_path)) {
save(ArgsList, file = file.path(metadata_path, paste0(ModelID, "_ArgsList.Rdata")))
}
}
# Data Prepare ----
if(DebugMode) print("Data Prepare ----")
Output <- H2ODataPrep(TargetType.="regression", TargetColumnName.=TargetColumnName, data.=data, ValidationData.=ValidationData, TestData.=TestData, TrainOnFull.=TrainOnFull, FeatureColNames.=FeatureColNames, SaveModelObjects.=SaveModelObjects, model_path.=metadata_path, ModelID.=ModelID, TransformNumericColumns.=TransformNumericColumns, Methods.=Methods)
TransformationResults <- Output$TransformationResults; Output$TransformationResults <- NULL
dataTrain <- Output$dataTrain; Output$dataTrain <- NULL
dataTest <- Output$dataTest; Output$dataTest <- NULL
TestData <- Output$TestData; Output$TestData <- NULL
MinVal <- Output$MinVal; Output$MinVal <- NULL
Names <- Output$Names; rm(Output)
# Grid Tune Check ----
if(GridTune && !TrainOnFull) {
# Grid tune ----
if(DebugMode) print("Grid tune ----")
# H2O init and data load
if(H2OStartUp) localHost <- h2o::h2o.init(nthreads = NThreads, max_mem_size = MaxMem, enable_assertions = FALSE)
datatrain <- h2o::as.h2o(dataTrain)
if(!TrainOnFull) datavalidate <- h2o::as.h2o(dataTest, use_datatable = TRUE) else datavalidate <- NULL
if(!is.null(TestData)) datatest <- h2o::as.h2o(TestData, use_datatable = TRUE) else datatest <- NULL
# Grid Tune Search Criteria ----
search_criteria <- list(
strategy = GridStrategy,
max_runtime_secs = MaxRunTimeSecs,
max_models = MaxModelsInGrid,
seed = 1234,
stopping_rounds = StoppingRounds,
stopping_metric = toupper(eval_metric),
stopping_tolerance = 1e-3)
# Grid Parameters ----
hyper_params <- list()
hyper_params[["alpha"]] <- Alpha
hyper_params[["lambda"]] <- Lambda
# Link ----
if(is.null(Link)) Link <- "identity"
# Grid Train Model ----
grid <- h2o::h2o.grid(
hyper_params = hyper_params,
search_criteria = search_criteria,
is_supervised = TRUE,
algorithm = "glm",
grid_id = paste0(ModelID, "_Grid"),
x = FeatureColNames,
y = TargetColumnName,
training_frame = datatrain,
validation_frame = datavalidate,
family = Distribution,
link = Link,
seed = 1234)
# Get Best Model ----
Grid_Out <- h2o::h2o.getGrid(grid_id = paste0(ModelID, "_Grid"), sort_by = eval_metric, decreasing = FALSE)
# Collect Best Grid Model ----
base_model <- h2o::h2o.getModel(Grid_Out@model_ids[[1L]])
}
# Start Up H2O ----
if(!GridTune) {
H2OArgs <- list()
if(!is.null(RandomColNumbers)) H2OArgs[["HGLM"]] <- TRUE
# Build Model ----
if(DebugMode) print("Build Model ----")
# Start Up H2O ----
if(H2OStartUp) localHost <- h2o::h2o.init(nthreads = NThreads, max_mem_size = MaxMem, enable_assertions = FALSE)
datatrain <- h2o::as.h2o(dataTrain)
if(!TrainOnFull) datavalidate <- h2o::as.h2o(dataTest, use_datatable = TRUE) else datavalidate <- NULL
if(!is.null(TestData) && !H2OArgs[['HGLM']]) datatest <- h2o::as.h2o(TestData, use_datatable = TRUE) else datatest <- NULL
# Define link ----
if(!GridTune) if(is.null(Link)) Link <- "identity"
# Define ml args ----
H2OArgs[["x"]] <- FeatureColNames
H2OArgs[["y"]] <- TargetColumnName
H2OArgs[["interactions"]] <- InteractionColNumbers
H2OArgs[["weights_column"]] <- WeightsColumn
H2OArgs[["training_frame"]] <- datatrain
H2OArgs[["family"]] <- Distribution
H2OArgs[["link"]] <- Link
H2OArgs[["model_id"]] <- ModelID
if(length(H2OArgs[['HGLM']]) > 0L) {
H2OArgs[["rand_family"]] <- RandomDistribution
H2OArgs[["rand_link"]] <- RandomLink
H2OArgs[["random_columns"]] <- RandomColNumbers
} else {
H2OArgs[["validation_frame"]] <- datavalidate
H2OArgs[["HGLM"]] <- FALSE
}
H2OArgs[["solver"]] <- Solver
H2OArgs[["alpha"]] <- Alpha
H2OArgs[["lambda"]] <- Lambda
H2OArgs[["lambda_search"]] <- LambdaSearch
H2OArgs[["nlambdas"]] <- NLambdas
H2OArgs[["standardize"]] <- Standardize
H2OArgs[["remove_collinear_columns"]] <- RemoveCollinearColumns
H2OArgs[["intercept"]] <- InterceptInclude
H2OArgs[["non_negative"]] <- NonNegativeCoefficients
H2OArgs[["tweedie_link_power"]] <- TweedieLinkPower
H2OArgs[["tweedie_variance_power"]] <- TweedieVariancePower
# Build model ----
base_model <- do.call(what = h2o::h2o.glm, args = H2OArgs)
}
# Save Final Model ----
if(DebugMode) print("Save Final Model ----")
H2OSaveModel(SaveModelObjects.=SaveModelObjects, IfSaveModel.=IfSaveModel, base_model.=base_model, model_path.=model_path, ModelID.=ModelID)
# Score Train Data ----
if(DebugMode) print("Score Final Test Data ----")
if("score_traindata" %chin% tolower(OutputSelection) && !TrainOnFull) {
Predict <- data.table::as.data.table(h2o::h2o.predict(object = base_model, newdata = datatrain))
}
# Create Train Validation Data ----
if(DebugMode) print("Create Validation Data ----")
if("score_traindata" %chin% tolower(OutputSelection) && !TrainOnFull) {
Output <- H2OValidationData(Predict.=Predict, TestData.=NULL, dataTest.=NULL, dataTrain.=dataTrain, TrainOnFull.=TRUE, SaveModelObjects.=SaveModelObjects, metadata_path.=metadata_path, model_path.=metadata_path, ModelID.=ModelID, TransformNumericColumns.=NULL, TransformationResults.=NULL, TargetColumnName.=NULL, data.=NULL)
TrainData <- Output$ValidationData; rm(Output)
}
# Score Validation Data
Predict <- data.table::as.data.table(h2o::h2o.predict(object = base_model, newdata = if(!is.null(TestData)) datatest else if(!TrainOnFull) datavalidate else datatrain))
# Create Validation Data ----
Output <- H2OValidationData(Predict.=Predict, TestData.=if(H2OArgs[['HGLM']]) NULL else TestData, dataTest.=dataTest, dataTrain.=dataTrain, TrainOnFull.=TrainOnFull, SaveModelObjects.=SaveModelObjects, metadata_path.=metadata_path, model_path.=metadata_path, ModelID.=ModelID, TransformNumericColumns.=NULL, TransformationResults.=NULL, TargetColumnName.=NULL, data.=NULL)
ValidationData <- Output$ValidationData; rm(Output)
# Variable Importance ----
if(DebugMode) print("Variable Importance ----")
VariableImportance <- H2OVariableImportance(TrainOnFull.=TrainOnFull, base_model.=base_model, SaveModelObjects.=SaveModelObjects, metadata_path.=metadata_path, model_path.=metadata_path, ModelID.=ModelID)
# H2O Explain TrainData ----
if(DebugMode) print("H2O Explain TrainData ----")
ExplainList <- list()
if(all(c("score_traindata","h2o.explain") %chin% tolower(OutputSelection)) && !TrainOnFull) {
ExplainList[["Train_Explain"]] <- h2o::h2o.explain(base_model, newdata = datatrain)
}
# H2O Explain ValidationData ----
if(DebugMode) print("H2O Explain ValidationData ----")
if(!TrainOnFull && "h2o.explain" %chin% tolower(OutputSelection)) {
ExplainList[["Test_Explain"]] <- h2o::h2o.explain(base_model, newdata = if(!is.null(TestData)) datatest else if(!is.null(ValidationData) && !TrainOnFull) datavalidate else datatrain)
}
# H2O Shutdown ----
if(DebugMode) print("H2O Shutdown ----")
if(H2OShutdown) h2o::h2o.shutdown(prompt = FALSE)
# Generate EvaluationMetrics ----
if(DebugMode) print("Generate EvaluationMetrics ----")
EvalMetricsList <- list()
if("evalmetrics" %chin% tolower(OutputSelection)) {
if("score_traindata" %chin% tolower(OutputSelection) && !TrainOnFull) {
EvalMetricsList[["TrainData"]] <- RegressionMetrics(SaveModelObjects.=SaveModelObjects, data.=data, ValidationData.=TrainData, TrainOnFull.=TrainOnFull, LossFunction.=eval_metric, EvalMetric.=eval_metric, TargetColumnName.=TargetColumnName, ModelID.=ModelID, model_path.=metadata_path, metadata_path.=metadata_path)
}
EvalMetricsList[["TestData"]] <- RegressionMetrics(SaveModelObjects.=SaveModelObjects, data.=data, ValidationData.=ValidationData, TrainOnFull.=TrainOnFull, LossFunction.=eval_metric, EvalMetric.=eval_metric, TargetColumnName.=TargetColumnName, ModelID.=ModelID, model_path.=metadata_path, metadata_path.=metadata_path)
}
# Subset Transformation Object ----
if(!is.null(TransformNumericColumns)) {
if(TargetColumnName == "Target") {
TransformationResults <- TransformationResults[!(ColumnName %chin% c("Predict"))]
} else {
TransformationResults <- TransformationResults[!(ColumnName %chin% c("Predict", "Target"))]
}
}
# Return Objects ----
if(DebugMode) print("Return Objects ----")
if(ReturnModelObjects) {
return(list(
Model = base_model,
TrainData = if(exists("TrainData") && !is.null(TrainData)) TrainData else NULL,
TestData = if(exists("ValidationData") && !is.null(ValidationData)) ValidationData else NULL,
H2OExplain = if(exists("ExplainList")) ExplainList else NULL,
EvaluationMetrics = if(exists("EvalMetricsList")) EvalMetricsList else NULL,
VariableImportance = if(exists("VariableImportance")) VariableImportance else NULL,
TransformationResults = if(exists("TransformationResults")) TransformationResults else NULL,
ColNames = if(exists("Names")) Names else NULL,
ArgsList = ArgsList))
}
}
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.