Improving a Regression Model: Reducing MAE with Ensemble Learning (RECS Dataset)

Описание: In this video, we present our regression modeling process for the RECS household electricity consumption dataset. Our goal was to predict average daily electricity usage while minimizing Mean Absolute Error (MAE).

We begin with simple baseline models and gradually improve performance through feature engineering, hyperparameter tuning, and model selection. We then transition to more advanced approaches, including HistGradientBoostingRegressor and CatBoost, and finally apply ensemble learning to combine model strengths.

To evaluate and improve model fit, we use:

Hold-out validation and K-Fold cross-validation

Out-of-Fold (OOF) error analysis

Learning curves to diagnose bias vs. variance

Clipping and log-transformation to stabilize predictions

These steps allowed us to systematically reduce MAE and achieve a significantly better leaderboard score. The video explains both what we changed and why those changes improved performance, focusing on practical machine learning workflow and model diagnostics rather than only final results.

This project demonstrates an end-to-end regression pipeline including preprocessing, feature engineering, model tuning, ensemble weighting, and performance validation.