Exploring The Variables Importance"

In ReSurv: Machine Learning Models for Predicting Claim Counts

knitr::opts_chunk$set(echo = TRUE)

library(ReSurv)

Introduction

Machine learning models catch interactions between covariates. Often they are a black-box but they can be interpreted with SHAP values. We generate two data sets, one from scenario Alpha and one from scenario Delta the plotting functionalities of the ReSurv package.

input_data_0 <- data_generator(
  random_seed = 1,
  scenario = 0,
  time_unit = 1 / 360,
  years = 4,
  yearly_exposure = 200
)

individual_data_0 <- IndividualDataPP(
  data = input_data_0,
  id = NULL,
  categorical_features = "claim_type",
  continuous_features = "AP",
  accident_period = "AP",
  calendar_period = "RP",
  input_time_granularity = "days",
  output_time_granularity = "quarters",
  years = 4
)

# Input data scenario Delta

input_data3 <- data_generator(
  random_seed = 1,
  scenario = 3,
  time_unit = 1 / 360,
  years = 4,
  yearly_exposure = 200
)

individual_data_3 <- IndividualDataPP(
  data = input_data3,
  id = NULL,
  categorical_features = "claim_type",
  continuous_features = "AP",
  accident_period = "AP",
  calendar_period = "RP",
  input_time_granularity = "days",
  output_time_granularity = "quarters",
  years = 4
)

Here we fit Neural Networks and XGB. In order to simplify this vignette, we provide in advance the optimal hyperparameters.

hp_scenario_alpha_xgb <- list(
  params = list(
    booster = "gbtree",
    eta = 0.9887265,
    subsample = 0.7924135 ,
    alpha = 10.85342,
    lambda = 6.213317,
    min_child_weight = 3.042204,
    max_depth = 1
  ),
  print_every_n = 0,
  nrounds = 3000,
  verbose = FALSE,
  early_stopping_rounds = 500
)

hp_scenario_alpha_nn <- list(
  batch_size = as.integer(5000),
  epochs = as.integer(5500),
  num_workers = 0,
  tie = 'Efron',
  num_layers = 2,
  num_nodes = 10,
  optim = "SGD",
  batch_size = as.integer(5000),
  lr = 0.3023043,
  xi = 0.426443,
  eps = 0,
  activation = "SELU",
  early_stopping = TRUE,
  patience = 350,
  verbose = FALSE,
  network_structure = NULL
)

hp_scenario_delta_xgb <- list(params=list(booster="gbtree",
                                          eta=0.2717736,
                                          subsample=0.9043068 ,
                                          alpha=7.789214,
                                          lambda=12.09398 ,
                                          min_child_weight=22.4837 ,
                                          max_depth = 4),
                                          print_every_n = 0,
                                          nrounds=3000,
                                          verbose= FALSE,
                                          early_stopping_rounds = 500)

hp_scenario_delta_nn <- list(
  batch_size = as.integer(5000),
  epochs = as.integer(5500),
  num_workers = 0,
  tie = 'Efron',
  num_layers = 2,
  num_nodes = 2,
  optim = "Adam",
  batch_size = as.integer(5000),
  lr = 0.3542422,
  xi = 0.1803953,
  eps = 0,
  activation = "LeakyReLU",
  early_stopping = TRUE,
  patience = 350,
  verbose = FALSE,
  network_structure = NULL
)

resurv_model_xgb_A <-  ReSurv(individual_data_0,
                              hazard_model = "XGB",
                              hparameters = hp_scenario_alpha_xgb)

resurv_model_nn_A <-  ReSurv(individual_data_0,
                             hazard_model = "NN",
                             hparameters = hp_scenario_alpha_nn)

resurv_model_xgb_D <-  ReSurv(individual_data_3,
                              hazard_model = "XGB",
                              hparameters = hp_scenario_delta_xgb)

resurv_model_nn_D <- ReSurv(individual_data_3,
                            hazard_model = "NN",
                            hparameters = hp_scenario_delta_nn)

Shap values (XGB)

plot(resurv_model_xgb_A)

plot(resurv_model_xgb_D)

Shap values (NN)

plot(resurv_model_nn_A, nsamples = 10000)

plot(resurv_model_nn_D, nsamples=10000)

ReSurv documentation built on April 4, 2025, 1:39 a.m.