Add Quantile Regression support to LightGbmRegressionTrainer

[ssperling-wmp]

Summary

The underlying LightGBM C++ framework natively supports quantile regression via the parameters objective="quantile" and alpha=<float> (where alpha is the target quantile, e.g., 0.05 for the 5th percentile). However, the ML.NET LightGbmRegressionTrainer wrapper currently hardcodes objective="regression" in LightGbmRegressionTrainer.cs and does not expose the alpha parameter, making it impossible to train quantile regression models through the ML.NET API.

Motivation

Quantile regression is widely used for:

• Prediction intervals: Training two models (e.g., α=0.05 and α=0.95) to obtain a 90% prediction interval
• Risk modeling: Predicting worst/best-case scenarios
• Heteroscedastic data: Where variance changes across the feature space

Since the native LightGBM library already supports this, exposing it through ML.NET would be a high-value, low-risk enhancement.

Proposed Changes

1. Add a RegressionObjective enum to LightGbmRegressionTrainer.Options with values Regression (default) and Quantile
2. Add an Alpha property (double, default 0.5) to LightGbmRegressionTrainer.Options
3. Modify CheckAndUpdateParametersBeforeTraining to conditionally set objective=quantile and alpha=<value>
4. Validate that Alpha ∈ (0, 1)
5. Add unit tests

API Usage

var options = new LightGbmRegressionTrainer.Options
{
    Objective = LightGbmRegressionTrainer.Options.RegressionObjective.Quantile,
    Alpha = 0.95,
};
var trainer = mlContext.Regression.Trainers.LightGbm(options);

Scope

• No new trainer class, quantile regression produces identical tree structures and output schema
• No changes to model serialization, ONNX export, or catalog extension methods

[github-actions]

🔍 No similar issues found with high confidence.

---

AI-powered similar issue detection