Credit Card

ML - Gradient Boosting (GBM)

The results from H2O point to using a Gradient Boosting (GBM) algorithm.

In this lab, you operationalize that choice in PDI:

• Train a GBM model in R.

• Save the model artifact.

• Predict fraudulent credit card transactions.

You will use the R gbm package.

Walkthrough (video)

Train the Model

Train a GBM model with the same dataset.

main_job.kjb

1. In Spoon, open the following main job:

~/Workshop--Data-Integration/Labs/Module 7 - Workflows/Machine Learning/Credit Card Fraud/solution/jb_fraud_main_job.kjb

1. Right-click the train_model transformation.

2. Select Open referenced object > Transformation.

train model

---

R Script Executor

1. Open the rscrpt-train_gbm step.

2. On the Configure tab, set:

   • Input frames: sv-convert_booleans_to_numbers

   • R frame name: train

1. Set Row handling > Number of rows to process to All.

2. On the R script tab, paste this script:

# ============================================================
# GBM Model Training - Credit Card Fraud Detection
# ============================================================
# This script trains a Gradient Boosting Machine (GBM) model
# to predict fraudulent credit card transactions.
# It runs inside the PDI R Script Executor step.
# ============================================================

# Load the GBM library for gradient boosting
library(gbm)

# Convert the incoming PDI data frame ("train") to a standard R data frame
# The "train" variable is automatically created by the R Script Executor
# from the input step: sv-convert_booleans_to_numbers
train.df <- as.data.frame(train)

# Convert the target variable to binary 0/1
# as.factor() creates levels, as.numeric() assigns 1 and 2, then subtract 1
# Result: 0 = not fraud, 1 = fraud
train.df$reported_as_fraud_historic <- as.numeric(
  as.factor(train.df$reported_as_fraud_historic)
) - 1

# Train the GBM model
# --------------------------------------------------------
# Note: We use OOB (out-of-bag) estimation instead of
# cross-validation (cv.folds) because JRI runs R inside
# the JVM process. Cross-validation spawns additional
# processes that exceed the FD_SETSIZE limit (1024) in
# Linux's select() system call, causing the JVM to abort.
# --------------------------------------------------------
gbm_model <- gbm(
  reported_as_fraud_historic ~ .,   # predict fraud using all other columns
  data = train.df,                  # training data
  distribution = "bernoulli",       # binary classification (fraud yes/no)
  n.trees = 500,                    # number of boosting iterations
  interaction.depth = 4,            # max depth of each tree
  shrinkage = 0.01,                 # learning rate (smaller = more robust)
  n.minobsinnode = 10,              # min observations per terminal node
  bag.fraction = 0.5                # use 50% of data per tree (stochastic GBM)
)

# Determine the optimal number of trees using OOB error
# This avoids overfitting by finding where performance plateaus
best_trees <- gbm.perf(gbm_model, method = "OOB")

# Save the trained model and optimal tree count to disk
# The predict transformation will load this file to score new transactions
save(gbm_model, best_trees,
  file = "/home/pentaho/Workshop--Data-Integration/Labs/Module 7 - Use Cases/Machine Learning/Credit Card Fraud/solution/train_model_output/gbm_fraud.rdata"
)

# Return a status message to PDI
# The R Script Executor expects a data frame as output
ok <- "Finished"
ok.df <- as.data.frame(ok)
ok.df
  )

  # Determine the optimal number of trees using Out-of-Bag (OOB) estimation.
  # gbm.perf() analyzes the OOB improvement curve and returns the iteration
  # (tree count) where the OOB error is minimized. Using more trees than this
  # would overfit; using fewer would underfit. This value will be used during
  # scoring to make predictions with only the best-performing subset of trees.
  best_trees <- gbm.perf(gbm_model, method = "OOB")

  # Save the trained model object and the optimal tree count to an .rdata file.
  # This file will be loaded later by a separate PDI scoring transformation
  # to apply the model to new/unseen transactions for real-time fraud detection.
  save(
    gbm_model,
    best_trees,
    file = "/home/pentaho/Workshop--Data-Integration/Labs/Module 7 - Use Cases/Machine Learning/Credit Card Fraud/solution/train_model_output/gbm_fraud.rdata"
  )

  # If we reach this point, training completed successfully.
  # This string is returned as the value of 'ok' by tryCatch().
  "Finished"

# Error handler: if ANY step above throws an R error, this function catches it
# and returns the error message as a string. This ensures the script always
# produces the 'ok' output column that PDI expects, while preserving the
# actual error details for troubleshooting in the PDI log.
}, error = function(e) {
  paste("ERROR:", e$message)
})

# Create a single-column data frame with the status result.
# The R Script Executor step in PDI expects a data frame as output.
# The column name 'ok' must match what is configured in the step's output fields.
#   - "Finished"       = training completed successfully
#   - "ERROR: <msg>"   = training failed; check the message for the root cause
# This allows downstream PDI steps (e.g., a Filter or Switch/Case) to route
# the flow based on success or failure of the model training.
ok.df <- as.data.frame(ok)
ok.df

This step writes the model artifact to:

/home/pentaho/Workshop--Data-Integration/Labs/Module 7 - Use Cases/Machine Learning/Credit Card Fraud/solution/train_model_output/gbm_fraud.rdata

Predict Fraud

Results