[Phase 4] Parameter Regression Models

[Sakeeb91]

Objective

Predict atmospheric parameters (Teff, log g, [Fe/H], [alpha/Fe]) with target MAE thresholds.

Dependencies

• Phase 2: Preprocessing Pipeline (requires cleaned features)
• Phase 3: Classification (can run in parallel)

Tasks

• Implement src/models/regressor.py with multi-output regression
• Implement src/evaluation/regression_metrics.py for MAE, R2, scatter
• Train separate models for each parameter
• Analyze residuals for systematic bias
• Create notebooks/03_regression_analysis.ipynb
• Save trained models to models/parameter_regressor_v1.joblib

Files to Create

File	Purpose
src/models/regressor.py	Regression models
src/evaluation/regression_metrics.py	Regression metrics
notebooks/03_regression_analysis.ipynb	Analysis

Target Metrics

Parameter	Target MAE
Teff	< 100 K
log g	< 0.2 dex
[Fe/H]	< 0.1 dex
[alpha/Fe]	< 0.05 dex

Starter Code

# src/models/regressor.py
"""Stellar parameter regression models."""

import numpy as np
import joblib
from sklearn.ensemble import RandomForestRegressor
from xgboost import XGBRegressor

class ParameterRegressor:
    """Regressor for stellar atmospheric parameters."""

    def __init__(self, target: str, model_type: str = "xgboost"):
        self.target = target
        self.model_type = model_type

        if model_type == "xgboost":
            self.model = XGBRegressor(
                n_estimators=100,
                max_depth=6,
                learning_rate=0.1,
                random_state=42
            )
        else:
            self.model = RandomForestRegressor(
                n_estimators=100,
                random_state=42
            )

    def fit(self, X: np.ndarray, y: np.ndarray) -> "ParameterRegressor":
        """Train the regressor."""
        self.model.fit(X, y)
        return self

    def predict(self, X: np.ndarray) -> np.ndarray:
        """Predict parameter values."""
        return self.model.predict(X)

    def save(self, path: str) -> None:
        """Save model to disk."""
        joblib.dump({"model": self.model, "target": self.target}, path)

Definition of Done

• Teff MAE < 100 K
• log g MAE < 0.2 dex
• [Fe/H] MAE < 0.1 dex
• Residual analysis shows no systematic bias
• All tests passing

Technical Notes

• Different parameters have different error characteristics
• Consider parameter correlations (Teff-log g relation)
• XGBoost often outperforms RF for regression

---

Part of #1 (Meta Issue)