Model calibration with pyspark

This package provides a Betacal class which allows the user to fit/train the default beta calibration model on pyspark dataframes and predict calibrated scores

Setup

spark-calibration package is uploaded to PyPi and can be installed with this command:

pip install spark-calibration

Usage

Training

train_df should be a pyspark dataframe containing:

• A column with raw model scores (default name: score)
• A column with binary labels (default name: label)

You can specify different column names when calling fit(). In some tree-based models like LightGBM, the predicted scores may fall outside the [0, 1] range and can even be negative. Please apply a sigmoid function to normalize the outputs accordingly.

from spark_calibration import Betacal
from spark_calibration import display_classification_calib_metrics
from spark_calibration import plot_calibration_curve

# Initialize model
bc = Betacal(parameters="abm")

# Load training data
train_df = spark.read.parquet("s3://train/")

# Fit the model
bc.fit(train_df)

# Or specify custom column names
# bc.fit(train_df, score_col="raw_score", label_col="actual_label")

# Access model parameters
print(f"Model coefficients: a={bc.a}, b={bc.b}, c={bc.c}")

The model learns three parameters:

• a: Coefficient for log(score)
• b: Coefficient for log(1-score)
• c: Intercept term

Saving and Loading Models

You can save the trained model to disk and load it later:

# Save model
save_path = bc.save("/path/to/save/")

# Load model
loaded_model = Betacal.load("/path/to/save/")

Prediction

test_df should be a pyspark dataframe containing a column with raw model scores. By default, this column should be named score, but you can specify a different column name when calling predict(). The predict function adds a new column prediction which has the calibrated score.

test_df = spark.read.parquet("s3://test/")

# Using default column name 'score'
test_df = bc.predict(test_df)

# Or specify a custom score column name
# test_df = bc.predict(test_df, score_col="raw_score")

Pre & Post Calibration Classification Metrics

The test_df should have score, prediction & label columns. The display_classification_calib_metrics functions displays brier_score_loss, log_loss, area_under_PR_curve and area_under_ROC_curve

display_classification_calib_metrics(test_df)

Output

model brier score loss: 0.08072683729933376
calibrated model brier score loss: 0.01014015353257748
delta: -87.44%

model log loss: 0.3038106859864252
calibrated model log loss: 0.053275633947890755
delta: -82.46%

model aucpr: 0.03471287564672635
calibrated model aucpr: 0.03471240518472563
delta: -0.0%

model roc_auc: 0.7490639506966398
calibrated model roc_auc: 0.7490649764289607
delta: 0.0%

Plot the Calibration Curve

Computes true, predicted probabilities (pre & post calibration) using quantile binning strategy with 50 bins and plots the calibration curve

plot_calibration_curve(test_df)