Rob Wiederstein
33d0fc31c7
End-to-end LightGBM fraud detection pipeline built as an R package, orchestrated by targets with data stored in MinIO via Apache Arrow. Includes 6-layer Lakehouse architecture, class imbalance tournament, formally tuned hyperparameters (PR-AUC 0.198), and Quarto RevealJS slides. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
327 lines
7.3 KiB
R
327 lines
7.3 KiB
R
library(targets)
|
|
library(tarchetypes)
|
|
|
|
tar_option_set(
|
|
packages = c(
|
|
"arrow",
|
|
"bonsai",
|
|
"duckdb",
|
|
"glue",
|
|
"gt",
|
|
"here",
|
|
"lightgbm",
|
|
"lubridate",
|
|
"tidymodels",
|
|
"tidyverse",
|
|
"cowplot",
|
|
"colorspace",
|
|
"readr",
|
|
"scales",
|
|
"ggplot2",
|
|
"quarto",
|
|
"corrr",
|
|
"recipes",
|
|
"themis",
|
|
"tidyselect"
|
|
)
|
|
)
|
|
|
|
tar_source("./R/functions.R")
|
|
|
|
list(
|
|
tar_target(
|
|
baf_parquet_prefix,
|
|
convert_to_parquet(
|
|
from_prefix = "01_raw",
|
|
to_prefix = "02_intermediate",
|
|
bucket_name = "baf-fraud"
|
|
)
|
|
),
|
|
|
|
tar_target(
|
|
baf_primary_prefix,
|
|
clean_baf_base(
|
|
in_prefix = baf_parquet_prefix,
|
|
out_prefix = "03_primary/variant=Base",
|
|
bucket_name = "baf-fraud",
|
|
partitioning = "month",
|
|
existing_data_behavior = "delete_matching",
|
|
verbose = TRUE
|
|
)
|
|
),
|
|
|
|
tar_target(
|
|
baf_feature_prefix,
|
|
engineer_features(
|
|
in_prefix = baf_primary_prefix,
|
|
out_prefix = "04_feature/variant=Base",
|
|
bucket_name = "baf-fraud",
|
|
partitioning = "month",
|
|
existing_data_behavior = "delete_matching",
|
|
verbose = TRUE
|
|
)
|
|
),
|
|
|
|
# ---- Figure objects ----
|
|
tar_target(
|
|
fig_fraud_by_month,
|
|
plot_fraud_by_month(baf_primary_prefix, bucket_name = "baf-fraud")
|
|
),
|
|
|
|
# ---- Saved figure path (file target) ----
|
|
tar_target(
|
|
fig_fraud_by_month_path,
|
|
save_report_figure(
|
|
fig_fraud_by_month,
|
|
filename = "fig_fraud_by_month.png",
|
|
out_dir = "reports/figures"
|
|
),
|
|
format = "file"
|
|
),
|
|
tar_target(
|
|
tbl_fraud_by_month_data,
|
|
compute_fraud_by_month(baf_primary_prefix)
|
|
),
|
|
|
|
tar_target(
|
|
tbl_fraud_by_month_gt,
|
|
format_fraud_by_month_gt(tbl_fraud_by_month_data)
|
|
),
|
|
|
|
tar_target(
|
|
tbl_fraud_by_month_path,
|
|
save_report_table(tbl_fraud_by_month_gt, filename = "tbl_fraud_by_month.rds"),
|
|
format = "file"
|
|
),
|
|
|
|
# ---- Exploratory Data Analysis (EDA) Layer ----
|
|
tar_target(
|
|
data_eda_m0,
|
|
connect_baf(baf_primary_prefix, use_duckdb = TRUE) |>
|
|
filter(month == 0) |>
|
|
collect()
|
|
),
|
|
|
|
tar_target(
|
|
eda_recipe,
|
|
prepare_eda_recipe(data_eda_m0)
|
|
),
|
|
|
|
tar_target(
|
|
data_baked_eda_m0,
|
|
bake(eda_recipe, new_data = data_eda_m0)
|
|
),
|
|
|
|
tar_target(
|
|
model_diag,
|
|
train_diag_model(data_baked_eda_m0)
|
|
),
|
|
|
|
# ---- EDA Figures ----
|
|
tar_target(fig_var_imp, plot_var_imp(model_diag)),
|
|
tar_target(fig_hexbin_interaction, plot_hexbin_interaction(data_baked_eda_m0)),
|
|
tar_target(fig_missingness, plot_missingness(data_eda_m0)),
|
|
tar_target(fig_num_cor, plot_num_cor(data_eda_m0)),
|
|
|
|
# ---- Saved EDA Figure Paths ----
|
|
tar_target(
|
|
fig_var_imp_path,
|
|
save_report_figure(fig_var_imp, "fig_var_imp.png"),
|
|
format = "file"
|
|
),
|
|
tar_target(
|
|
fig_hexbin_interaction_path,
|
|
save_report_figure(fig_hexbin_interaction, "fig_hexbin_interaction.png"),
|
|
format = "file"
|
|
),
|
|
tar_target(
|
|
fig_missingness_path,
|
|
save_report_figure(fig_missingness, "fig_missingness.png"),
|
|
format = "file"
|
|
),
|
|
tar_target(
|
|
fig_num_cor_path,
|
|
save_report_figure(fig_num_cor, "fig_num_cor.png"),
|
|
format = "file"
|
|
),
|
|
|
|
# ---- 05_model_input Generation ----
|
|
tar_target(
|
|
model_inputs_prefix,
|
|
generate_model_inputs(
|
|
feature_prefix = baf_feature_prefix,
|
|
out_prefix = "05_model_input",
|
|
bucket_name = "baf-fraud"
|
|
)
|
|
),
|
|
|
|
# ---- Tournament Inputs ----
|
|
tar_target(
|
|
imbalance_tasks,
|
|
tibble::tribble(
|
|
~recipe_name, ~data_folder, ~scale_pos_weight,
|
|
"Standard", "baseline", 1,
|
|
"Weighted", "baseline", 4,
|
|
"Under", "under", 1,
|
|
"Smote", "smote", 1,
|
|
"Adasyn", "adasyn", 1,
|
|
"Tomek", "tomek", 1
|
|
)
|
|
),
|
|
|
|
tar_target(
|
|
imbalance_windows,
|
|
tibble::tribble(
|
|
~window_id, ~train_months, ~test_month,
|
|
"Window 1", c(0, 1, 2), 3,
|
|
"Window 2", c(1, 2, 3), 4,
|
|
"Window 3", c(2, 3, 4), 5
|
|
)
|
|
),
|
|
|
|
# ---- 1. Data Layer (The Tournament Results) ----
|
|
tar_target(
|
|
tbl_strategy_showdown,
|
|
{
|
|
# Force DAG to wait for the folders to be generated
|
|
force(model_inputs_prefix)
|
|
# Pass baf_feature_prefix so it tracks the latest layer
|
|
run_imbalance_tournament(imbalance_tasks, imbalance_windows, baf_feature_prefix)
|
|
}
|
|
),
|
|
|
|
# ---- 2. Figure Layer ----
|
|
tar_target(
|
|
fig_strategy_showdown,
|
|
create_efficiency_plot(tbl_strategy_showdown)
|
|
),
|
|
|
|
tar_target(
|
|
fig_strategy_showdown_path,
|
|
save_report_figure(
|
|
fig_strategy_showdown,
|
|
filename = "fig_strategy_showdown.png",
|
|
out_dir = "reports/figures"
|
|
),
|
|
format = "file"
|
|
),
|
|
|
|
# ---- 3. Table Layer (gt object) ----
|
|
tar_target(
|
|
tbl_strategy_showdown_gt,
|
|
format_class_imbalance_tourney_gt(tbl_strategy_showdown)
|
|
),
|
|
|
|
tar_target(
|
|
tbl_strategy_showdown_path,
|
|
save_report_table(
|
|
tbl_strategy_showdown_gt,
|
|
filename = "tbl_strategy_showdown.rds",
|
|
out_dir = "reports/tables"
|
|
),
|
|
format = "file"
|
|
),
|
|
|
|
# ---- Final Production Evaluation ----
|
|
tar_target(
|
|
final_eval_data,
|
|
evaluate_final_model(params = winning_params)
|
|
),
|
|
|
|
tar_target(
|
|
final_conf_mat,
|
|
yardstick::conf_mat(final_eval_data, truth, pred_class)
|
|
),
|
|
|
|
tar_target(
|
|
final_roc_curve,
|
|
yardstick::roc_curve(final_eval_data, truth, prob)
|
|
),
|
|
|
|
tar_target(
|
|
final_pr_curve,
|
|
yardstick::pr_curve(final_eval_data, truth, prob)
|
|
),
|
|
|
|
# ---- Save Final Assets ----
|
|
tar_target(
|
|
fig_final_curves,
|
|
{
|
|
p1 <- ggplot2::autoplot(final_roc_curve) + ggplot2::labs(title = "ROC Curve (Months 6-7)")
|
|
p2 <- ggplot2::autoplot(final_pr_curve) + ggplot2::labs(title = "PR Curve (Months 6-7)")
|
|
cowplot::plot_grid(p1, p2)
|
|
}
|
|
),
|
|
|
|
tar_target(
|
|
fig_final_curves_path,
|
|
save_report_figure(fig_final_curves, "fig_final_curves.png"),
|
|
format = "file"
|
|
),
|
|
|
|
tar_target(
|
|
tbl_final_conf_mat_path,
|
|
save_report_table(final_conf_mat, "tbl_final_conf_mat.rds", out_dir = "reports/tables"),
|
|
format = "file"
|
|
),
|
|
# ---- Generate and Save Heatmap ----
|
|
tar_target(
|
|
fig_final_conf_mat,
|
|
plot_conf_mat_heatmap(final_conf_mat)
|
|
),
|
|
|
|
tar_target(
|
|
fig_final_conf_mat_path,
|
|
save_report_figure(fig_final_conf_mat, "fig_final_conf_mat.png"),
|
|
format = "file"
|
|
),
|
|
# ---- Report Dependency Update ----
|
|
tar_target(
|
|
report_assets,
|
|
c(
|
|
fig_fraud_by_month_path,
|
|
tbl_fraud_by_month_path,
|
|
fig_strategy_showdown_path,
|
|
tbl_strategy_showdown_path,
|
|
fig_var_imp_path,
|
|
fig_hexbin_interaction_path,
|
|
fig_missingness_path,
|
|
fig_num_cor_path
|
|
),
|
|
format = "file"
|
|
),
|
|
|
|
tar_quarto(
|
|
report_slides,
|
|
path = "index.qmd"
|
|
),
|
|
# production model deployment
|
|
tar_target(
|
|
data_full,
|
|
connect_baf(baf_feature_prefix, use_duckdb = TRUE) |>
|
|
collect()
|
|
),
|
|
tar_target(
|
|
production_recipe_blueprint,
|
|
build_baf_recipe(data_full)
|
|
),
|
|
tar_target(
|
|
winning_params,
|
|
list(
|
|
trees = 844,
|
|
tree_depth = 3,
|
|
learn_rate = 0.0204,
|
|
min_n = 389
|
|
)
|
|
),
|
|
tar_target(
|
|
production_model_uri,
|
|
train_production_model(
|
|
data = data_full,
|
|
recipe = production_recipe_blueprint, # <--- Pass the untrained blueprint!
|
|
best_params = winning_params,
|
|
model_filename = "baf_lgbm_prod_v1.txt"
|
|
),
|
|
format = "rds"
|
|
)
|
|
) |