Density Plot of Beta Corrections for a Variable

Description

Generates a density plot showing the distribution of corrected Beta values to a GLM coefficient, along with the original Beta coefficient, and standard error bounds around it.

NOTE This function signature documents the interface of functions created by create_beta_corrected_density.

Usage

beta_corrected_density(varname, q = 0.05, type = "kde")

Arguments

Details

The plot shows:

• Density curve of corrected coefficient values

• Solid vertical line at the original GLM coefficient

• Dashed lines at plus/minus 1 standard error from the coefficient

• Automatic x-axis limits that cut off the highest and lowest q pct. If you want axis unaltered, set q = 0

Value

ggplot object(s) showing the density distribution of corrected beta coefficients with vertical lines indicating the original coefficient value and standard error bounds.

The item returned will be:

• single ggplot object when 'varname' was a numerical variable OR a coefficient name

• list of ggplot objects when 'varname' was a categorical variable

See Also

create_beta_corrected_density, explain_iblm

Examples

# This function is created inside explain_iblm() and is output as an item

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

explain_objects <- explain_iblm(iblm_model, df_list$test)

# plot can be for a categorical variable (produces list of plots, one for each level)
explain_objects$beta_corrected_density(varname = "Area")

# plot can be for a single categorical level
explain_objects$beta_corrected_density(varname = "AreaB")

# output can be numerical variable
explain_objects$beta_corrected_density(varname = "DrivAge")


# This function must be created, and cannot be called directly from the package
try(
beta_corrected_density(varname = "DrivAge")
)

Scatter Plot of Beta Corrections for a Variable

Description

Generates a scatter plot or boxplot showing SHAP corrections for a specified variable from a fitted model. For numerical variables, creates a scatter plot with optional coloring and marginal densities. For categorical variables, creates a boxplot with model coefficients overlaid.

NOTE This function signature documents the interface of functions created by create_beta_corrected_scatter.

Usage

beta_corrected_scatter(varname, q = 0, color = NULL, marginal = FALSE)

Arguments

Details

The function handles both numerical and categorical variables differently:

• Numerical: Creates scatter plot of variable values vs. beta + SHAP deviations

• Categorical: Creates boxplot of SHAP deviations for each level with coefficient overlay

For numerical variables, horizontal lines show the model coefficient (solid) and confidence intervals (dashed). SHAP corrections represent local deviations from the global model coefficient.

Value

A ggplot2 object. For numerical variables: scatter plot with SHAP corrections, model coefficient line, and confidence bands. For categorical variables: boxplot with coefficient points overlaid.

See Also

create_beta_corrected_scatter, explain_iblm

Examples

# This function is created inside explain_iblm() and is output as an item

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

explain_objects <- explain_iblm(iblm_model, df_list$test)

# plot can be for a categorical variable (box plot)
explain_objects$beta_corrected_scatter(varname = "Area")

# plot can be for a numerical variable (scatter plot)
explain_objects$beta_corrected_scatter(varname = "DrivAge")


# This function must be created, and cannot be called directly from the package
try(
beta_corrected_scatter(varname = "DrivAge")
)

Compute Beta Corrections based on SHAP values

Description

Processes SHAP values in one-hot (wide) format to create beta coefficient corrections.

This includes:

• scaling shap values of continuous variables by the predictor value for that row

• migrating shap values to the bias for continuous variables where the predictor value was zero

• migrating shap values to the bias for categorical variables where the predictor value was reference level

Usage

beta_corrections_derive(
  shap_wide,
  wide_input_frame,
  iblm_model,
  migrate_reference_to_bias = TRUE
)

Arguments

Value

A data frame with the booster model beta corrections in one-hot (wide) format

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

shap <- stats::predict(
  iblm_model$booster_model,
  newdata = xgboost::xgb.DMatrix(
    data.matrix(
      dplyr::select(df_list$test, -dplyr::all_of(iblm_model$response_var))
    )
  ),
  predcontrib = TRUE
) |> data.frame()

wide_input_frame <- data_to_onehot(df_list$test, iblm_model)

shap_wide <- shap_to_onehot(shap, wide_input_frame, iblm_model)

beta_corrections <- beta_corrections_derive(shap_wide, wide_input_frame, iblm_model)

beta_corrections |> dplyr::glimpse()

Density Plot of Bias Corrections from SHAP values

Description

Visualizes the distribution of SHAP corrections that are migrated to bias terms, showing both per-variable and total bias corrections.

NOTE This function signature documents the interface of functions created by create_bias_density.

Usage

bias_density(q = 0, type = "hist")

Arguments

Value

A list with two ggplot objects:

• bias_correction_var: Faceted plot showing bias correction density from each variable. Note that variables with no records contributing to bias correction are dropped from the plot.

• bias_correction_total: Plot showing total corrected bias density.

See Also

create_bias_density, explain_iblm

Examples

# This function is created inside explain_iblm() and is output as an item

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

explain_objects <- explain_iblm(iblm_model, df_list$test)

explain_objects$bias_density()


# This function must be created, and cannot be called directly from the package
try(
bias_density()
)

Check Data Variability for Modeling

Description

Validates that the response variable and all predictor variables have more than one unique value.

Usage

check_data_variability(data, response_var)

Arguments

Value

Invisibly returns 'TRUE' if all checks pass, otherwise throws an error.

Check Object of Class 'iblm'

Description

Validates an iblm model object has required structure and features

Usage

check_iblm_model(model, booster_models_supported = c("xgb.Booster"))

Arguments

Value

Invisible TRUE if all checks pass

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

check_iblm_model(iblm_model)

Plot GLM vs IBLM Predictions with Different Corridors

Description

Creates a faceted scatter plot comparing GLM predictions to ensemble predictions across different trim values, showing how the ensemble corrects the base GLM model.

Usage

correction_corridor(
  iblm_model,
  data,
  trim_vals = c(NA_real_, 4, 1, 0.2, 0.1, 0),
  sample_perc = 0.2,
  color = NA,
  ...
)

Arguments

Value

A ggplot object showing GLM vs IBLM predictions faceted by trim value. The diagonal line (y = x) represents perfect agreement between models

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

correction_corridor(iblm_model = iblm_model, data = df_list$test, color = "DrivAge")

Create Pre-Configured Beta Corrected Density Plot Function

Description

Factory function that returns a plotting function with data pre-configured.

Usage

create_beta_corrected_density(
  wide_input_frame,
  beta_corrections,
  data,
  iblm_model
)

Arguments

Value

Function with signature function(varname, q = 0.05, type = "kde").

See Also

[beta_corrected_density()]

Examples

# ------- prepare iblm objects required -------

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

test_data <- df_list$test
shap <- extract_booster_shap(iblm_model$booster_model, test_data)
wide_input_frame <- data_to_onehot(test_data, iblm_model)
shap_wide <- shap_to_onehot(shap, wide_input_frame, iblm_model)
beta_corrections <- beta_corrections_derive(shap_wide, wide_input_frame, iblm_model)

# ------- demonstration of functionality -------

# create_beta_corrected_density() can create function of type 'beta_corrected_density'
my_beta_corrected_density <- create_beta_corrected_density(
  wide_input_frame,
  beta_corrections,
  test_data,
  iblm_model
)

# this custom function then acts as per beta_corrected_density()
my_beta_corrected_density(varname = "VehAge")

Create Pre-Configured Beta Corrected Scatter Plot Function

Description

Factory function that returns a plotting function with data pre-configured.

Usage

create_beta_corrected_scatter(data_beta_coeff, data, iblm_model)

Arguments

Value

Function with signature function(varname, q = 0, color = NULL, marginal = FALSE).

See Also

[beta_corrected_scatter()]

Examples

# ------- prepare iblm objects required -------

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

test_data <- df_list$test
shap <- extract_booster_shap(iblm_model$booster_model, test_data)
wide_input_frame <- data_to_onehot(test_data, iblm_model)
shap_wide <- shap_to_onehot(shap, wide_input_frame, iblm_model)
beta_corrections <- beta_corrections_derive(shap_wide, wide_input_frame, iblm_model)
data_glm <- data_beta_coeff_glm(test_data, iblm_model)
data_booster <- data_beta_coeff_booster(test_data, beta_corrections, iblm_model)
data_beta_coeff <- data_glm + data_booster

# ------- demonstration of functionality -------

# create_beta_corrected_scatter() can create function of type 'beta_corrected_scatter'
my_beta_corrected_scatter <- create_beta_corrected_scatter(data_beta_coeff, test_data, iblm_model)

# this custom function then acts as per beta_corrected_scatter()
my_beta_corrected_scatter(varname = "VehAge")

Create Pre-Configured Bias Density Plot Function

Description

Factory function that returns a plotting function with data pre-configured.

Usage

create_bias_density(shap, data, iblm_model, migrate_reference_to_bias = TRUE)

Arguments

Value

Function with signature function(q = 0, type = "hist").

See Also

[bias_density()]

Examples

# ------- prepare iblm objects required -------

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

test_data <- df_list$test
shap <- extract_booster_shap(iblm_model$booster_model, test_data)

# ------- demonstration of functionality -------

# create_bias_density() can create function of type 'bias_density'
my_bias_density <- create_bias_density(shap, test_data, iblm_model)

# this custom function then acts as per bias_density()
my_bias_density()

Create Pre-Configured Overall Correction Plot Function

Description

Factory function that returns a plotting function with data pre-configured.

Usage

create_overall_correction(shap, iblm_model)

Arguments

Value

Function with signature function(transform_x_scale_by_link = TRUE).

See Also

[overall_correction()]

Examples

# ------- prepare iblm objects required -------

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

test_data <- df_list$test
shap <- extract_booster_shap(iblm_model$booster_model, test_data)

# ------- demonstration of functionality -------

# create_overall_correction() can create function of type 'overall_correction'
my_overall_correction <- create_overall_correction(shap, iblm_model)

# this custom function then acts as per overall_correction()
my_overall_correction()

Obtain Booster Model Beta Corrections for tabular data

Description

Creates dataframe of Shap beta corrections for each row and predictor variable of 'data'

Usage

data_beta_coeff_booster(data, beta_corrections, iblm_model)

Arguments

Value

A data frame with beta coefficient corrections. The structure will be the same dimension as 'data' except for a "bias" column at the start.

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

explainer_outputs <- explain_iblm(iblm_model, df_list$test)

data_booster <- data_beta_coeff_booster(
  df_list$test,
  explainer_outputs$beta_corrections,
  iblm_model
)

data_booster |> dplyr::glimpse()

Obtain GLM Beta Coefficients for tabular data

Description

Creates dataframe of GLM beta coefficients for each row and predictor variable of 'data'

Usage

data_beta_coeff_glm(data, iblm_model)

Arguments

Value

A data frame with beta coefficients. The structure will be the same dimension as 'data' except for a "bias" column at the start.

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

data_glm <- data_beta_coeff_glm(df_list$train, iblm_model)

data_glm |> dplyr::glimpse()

Convert Data Frame to Wide One-Hot Encoded Format

Description

Transforms categorical variables in a data frame into one-hot encoded format

Usage

data_to_onehot(data, iblm_model, remove_target = TRUE)

Arguments

Value

A data frame in wide format with one-hot encoded categorical variables, an intercept column, and all variables ordered according to "coeff_names$all" from 'iblm_model'

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

wide_input_frame <- data_to_onehot(df_list$test, iblm_model)

wide_input_frame |> dplyr::glimpse()

Explain GLM Model Predictions Using SHAP Values

Description

Creates a list that explains the beta values, and their corrections, of the ensemble IBLM model

Usage

explain_iblm(iblm_model, data, migrate_reference_to_bias = TRUE)

Arguments

Details

The following outputs are functions that can be called to create plots:

• beta_corrected_scatter

• beta_corrected_density

• bias_density

• overall_correction

For each of these, the key data arguments (e.g. data, shap, iblm_model) are already populated by 'explain_iblm()'. When calling these functions output from 'explain_iblm()' only key settings like variable names, colours...etc need populating.

Value

A list containing:

beta_corrected_scatter

Function to create scatter plots showing SHAP corrections vs variable values (see beta_corrected_scatter)

beta_corrected_density

Function to create density plots of SHAP corrections for variables (see beta_corrected_density)

bias_density

Function to create density plots of SHAP corrections migrated to bias (see bias_density)

overall_correction

Function to show global correction distributions (see overall_correction)

shap

Dataframe showing raw SHAP values of data records

beta_corrections

Dataframe showing beta corrections (in wide/one-hot format) of data records

data_beta_coeff

Dataframe showing beta coefficients of data records

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

ex <- explain_iblm(iblm_model, df_list$test)

# the output contains functions that can be called to visualise iblm
ex$beta_corrected_scatter("DrivAge")
ex$beta_corrected_density("DrivAge")
ex$overall_correction()
ex$bias_density()

# the output contains also dataframes
ex$shap |> dplyr::glimpse()
ex$beta_corrections |> dplyr::glimpse()
ex$data_beta_coeff |> dplyr::glimpse()

Extract SHAP values from an xgboost Booster model

Description

A function to extract SHAP (SHapley Additive exPlanations) values from fitted booster model

Usage

extract_booster_shap(booster_model, data, ...)

## S3 method for class 'xgb.Booster'
extract_booster_shap(booster_model, data, ...)

Arguments

Details

Currently only a booster_model of class 'xgb.Booster' is supported

Value

A data frame of SHAP values, where each column corresponds to a feature and each row corresponds to an observation.

Methods (by class)

• extract_booster_shap(xgb.Booster): Extract SHAP values from an 'xgb.Booster' model

See Also

[xgboost::predict.xgb.Booster()]

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

booster_shap <- extract_booster_shap(iblm_model$booster_model, df_list$test)

booster_shap |> dplyr::glimpse()

French Motor Insurance Claims Dataset

Description

A dataset containing information about French motor insurance policies and claims, commonly used for actuarial modeling and risk assessment studies.

This is a "mini" subset of the CASdatasets 'freMTPL2freq' data, with some manipulation (see details)

Usage

freMTPLmini

Format

A data frame with 25,000 rows and 8 variables:

ClaimRate

Number of claims made, at an annualised rate, rounded (integer)

VehPower

Vehicle power rating or engine horsepower category (integer)

VehAge

Age of the vehicle in years (integer)

DrivAge

Age of the driver in years (integer)

BonusMalus

Bonus-malus coefficient, a rating factor used in French insurance where lower values indicate better driving records (integer)

VehBrand

Vehicle brand/manufacturer code (factor with levels like B6, B12, etc.)

VehGas

Type of fuel used by the vehicle (factor with levels: Regular, Diesel)

Area

Area classification where the policy holder resides (factor with levels A through F)

Details

The dataset is a random sample of 50,000 records from 'freMTPL2freq' from the 'CASdatasets' pacakge. Other modifications applied are:

• ClaimRate: Converted to ClaimNb per Exposure, winsorized at the 99.9th percentile, and rounded.

• VehAge: Ceiling of 50 years applied

• Dropped columns: Region, Density, Exposure, ClaimNb, IDpol

Source

['https://github.com/dutangc/CASdatasets/raw/c49cbbb37235fc49616cac8ccac32e1491cdc619/data/freMTPL2freq.rda']

Examples

head(freMTPLmini)

Calculate Pinball Scores for IBLM and Additional Models

Description

Computes Poisson deviance and pinball scores for an IBLM model alongside homogeneous, GLM, and optional additional models.

Usage

get_pinball_scores(
  data,
  iblm_model,
  trim = NA_real_,
  additional_models = list()
)

Arguments

Details

Pinball scores are calculated relative to a homogeneous model (i.e. a simple mean prediction of training data). Higher scores indicate better predictive performance.

Value

Data frame with 3 columns:

• "model" - will be homog, glm, iblm and any other models specified in 'additional_models'

• "'family'_deviance" - the value from the loss function based on the family of the glm function

• "pinball_score" - The more positive the score, the better the model than a basic homog model (i.e. all predictions are mean value). A negative score indicates worse than homog model.

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

get_pinball_scores(data = df_list$test, iblm_model = iblm_model)

Load French Motor Third-Party Liability Frequency Dataset

Description

Downloads the French Motor Third-Party Liability (freMTPL2freq) dataset from the CASdatasets GitHub repository, and apply minor transformations.

Usage

load_freMTPL2freq()

Details

The function performs the following modifications to the sourced:

• ClaimNb: Converted to ClaimNb per Exposure, and winsorized at the 99.9th percentile

• VehAge: Ceiling of 50 years applied

• All factor variables converted to character type

Value

A data frame containing the following columns:

ClaimNb

Number of claims per year.

VehPower

The power of the car (ordered categorical).

VehAge

The vehicle age in years, capped at 50.

DrivAge

The driver age in years (minimum 18, the legal driving age in France).

BonusMalus

Bonus/malus coefficient, ranging from 50 to 350. Values below 100 indicate a bonus (discount), while values above 100 indicate a malus (surcharge) in the French insurance system.

VehBrand

The car brand (categorical, with unknown/proprietary category labels).

VehGas

The car fuel type: "Diesel" or "Regular".

Area

The density classification of the city community where the driver lives, ranging from "A" (rural area) to "F" (urban centre).

Density

The population density (inhabitants per square kilometer) of the city where the driver lives.

Region

The policy region in France, based on the 1970-2015 administrative classification.

Note

This function requires an internet connection to download the data.

References

Dutang, C. CASdatasets: Insurance datasets. https://github.com/dutangc/CASdatasets/raw/c49cbbb37235fc49616cac8ccac32e1491cdc619/data/freMTPL2freq.rda

Examples

# Load the preprocessed dataset
freMTPL2freq <- load_freMTPL2freq()

freMTPL2freq |> dplyr::glimpse()

Plot Overall Corrections from Booster Component

Description

Creates a visualization showing for each record the overall booster component (either multiplicative or additive)

NOTE This function signature documents the interface of functions created by create_overall_correction.

Usage

overall_correction(transform_x_scale_by_link = TRUE)

Arguments

Value

A ggplot2 object.

See Also

create_overall_correction, explain_iblm

Examples

# This function is created inside explain_iblm() and is output as an item

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

explain_objects <- explain_iblm(iblm_model, df_list$test)

explain_objects$overall_correction()


# This function must be created, and cannot be called directly from the package
try(
overall_correction()
)

Predict Method for IBLM

Description

This function generates predictions from an ensemble model consisting of a GLM and an XGBoost model.

Usage

## S3 method for class 'iblm'
predict(object, newdata, trim = NA_real_, type = "response", ...)

Arguments

Details

The prediction process involves the following steps:

1. Generate GLM predictions

2. Generate Booster predictions

3. If trimming is specified, apply to booster predictions

4. Combine GLM and Booster predictions as per "relationship" described within iblm model object

At this point, only an iblm model with a "booster_model" object of class 'xgb.Booster' is supported

Value

A numeric vector of ensemble predictions computed as the element-wise product of GLM response probabilities and (optionally trimmed) XGBoost predictions.

See Also

predict.glm, predict.xgb.Booster

Examples

data <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  data,
  response_var = "ClaimRate",
  family = "poisson"
)

predictions <- predict(iblm_model, data$test)

predictions |> dplyr::glimpse()

Convert Shap values to Wide One-Hot Encoded Format

Description

Transforms categorical variables in a data frame into one-hot encoded format. Renames "BIAS" to lowercase.

Usage

shap_to_onehot(shap, wide_input_frame, iblm_model)

Arguments

Value

A data frame where SHAP values are in wide format for categorical variables. Column "bias" is moved to start.

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

shap <- stats::predict(
  iblm_model$booster_model,
  newdata = xgboost::xgb.DMatrix(
    data.matrix(
      dplyr::select(df_list$test, -dplyr::all_of(iblm_model$response_var))
    )
  ),
  predcontrib = TRUE
) |> data.frame()

wide_input_frame <- data_to_onehot(df_list$test, iblm_model)

shap_wide <- shap_to_onehot(shap, wide_input_frame, iblm_model)

shap_wide |> dplyr::glimpse()

Split Dataframe into: 'train', 'validate', 'test'

Description

This function randomly splits a data frame into three subsets for machine learning workflows: training, validation, and test sets. The proportions can be customized and must sum to 1.

Usage

split_into_train_validate_test(
  df,
  train_prop = 0.7,
  validate_prop = 0.15,
  test_prop = 0.15,
  seed = NULL
)

Arguments

Details

The function assigns each row to either "train", "validate" or "test" with the probability defined in the function.

Because each row is assigned a bucket independently, for very small datasets the proportions may not be as desired. This should not be an issue as data used for 'iblm' must be reasonably large.

Value

A named list with three elements:

train

A data frame containing the training subset

validate

A data frame containing the validation subset

test

A data frame containing the test subset

Examples

# Using 'mtcars'
split_into_train_validate_test(
  mtcars,
  train_prop = 0.6,
  validate_prop = 0.2,
  test_prop = 0.2,
  seed = 9000
)

Custom ggplot2 Theme for IBLM

Description

Custom ggplot2 Theme for IBLM

Usage

theme_iblm()

Value

A ggplot2 theme object that can be added to plots.

Train IBLM Model on XGBoost

Description

This function trains an interpretable boosted linear model.

The function combines a Generalized Linear Model (GLM) with a booster model of XGBoost

The "booster" model is trained on: - actual responses / GLM predictions, when the link function is log - actual responses - GLM predictions, when the link function is identity

Usage

train_iblm_xgb(
  df_list,
  response_var,
  family = "poisson",
  params = list(),
  nrounds = 1000,
  obj = NULL,
  feval = NULL,
  verbose = 0,
  print_every_n = 1L,
  early_stopping_rounds = 25,
  maximize = NULL,
  save_period = NULL,
  save_name = "xgboost.model",
  xgb_model = NULL,
  callbacks = list(),
  ...,
  strip_glm = TRUE
)

Arguments

Details

The 'family' argument will be fed into the GLM fitting. Default values for the XGBoost fitting are also selected based on family.

Note: Any xgboost configuration below will be overwritten by any explicit arguments input via 'params'

For "poisson" family the link function is 'log' and XGBoost is configured with:

• objective: "count:poisson"

• base_score: 1

For "gamma" family the link function is 'log' and XGBoost is configured with:

• objective: "reg:gamma"

• base_score: 1

For "tweedie" family the link function is 'log' (with a var.power = 1.5) and XGBoost is configured with:

• objective: "reg:tweedie"

• base_score: 1

• tweedie_variance_power = 1.5

For "gaussian" family the link function is 'identity' and XGBoost is configured with:

• objective: "reg:squarederror"

• base_score: 0

Value

An object of class "iblm" containing:

See Also

glm, xgb.train

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

iblm_model <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson"
)

Train XGBoost Model Using the IBLM Model Parameters

Description

Trains an XGBoost model using parameters extracted from the booster residual component of the iblm model. This is a convenient way to fit an XGBoost model for direct comparison with a fitted iblm model

Usage

train_xgb_as_per_iblm(iblm_model, ...)

Arguments

Value

Trained XGBoost model object (class "xgb.Booster").

See Also

xgb.train

Examples

df_list <- freMTPLmini |> split_into_train_validate_test(seed = 9000)

# training with plenty of rounds allowed
iblm_model1 <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson",
  max_depth = 6,
  nrounds = 1000
)

xgb1 <- train_xgb_as_per_iblm(iblm_model1)

# training with severe restrictions (expected poorer results)
iblm_model2 <- train_iblm_xgb(
  df_list,
  response_var = "ClaimRate",
  family = "poisson",
  max_depth = 1,
  nrounds = 5
)

xgb2 <- train_xgb_as_per_iblm(iblm_model2)

# comparison shows the poor training mirrored in second set:
get_pinball_scores(
  df_list$test,
  iblm_model1,
  trim = NA_real_,
  additional_models = list(iblm2 = iblm_model2, xgb1 = xgb1, xgb2 = xgb2)
)