Machine Learning Project
Hybrid ML for Quality Prediction
Feb 2024 – Jun 2024
A robust prediction pipeline combining PCA, Huber loss, classical models, and ensemble learning.
Modeling challenge
Wine-quality prediction involves correlated physicochemical measurements, noisy labels, and residuals that can be dominated by a small number of atypical samples. A model that ignores those properties can fit unstable directions in the data.
Dimensionality and signal
Principal component analysis was used as targeted denoising rather than generic compression. It reduced redundancy among correlated measurements and exposed lower-dimensional structure that could be used by downstream predictors.
Robust optimization
Huber loss interpolates between squared and absolute error. Small residuals retain the smooth behavior of an L2 objective, while large residuals receive less leverage, improving stability when labels are noisy or heavy-tailed.
Hybrid pipeline
The transformed representation and robust objective were evaluated with classical linear models and ensemble methods such as XGBoost. This design separates data conditioning, optimization robustness, and predictive capacity so that each component can be tested independently.