RFF Wine Quality Classifier

A fully functional PyTorch machine learning pipeline that applies Random Fourier Features to white wine quality prediction. The project uses a fixed random Fourier mapping as a kernel-inspired feature layer, then trains a small neural network classifier on top of the transformed tabular features.

The main experiment maps wine quality scores into three classes and compares a baseline RFF model against four tuned variants. A second binary experiment evaluates the same tuned RFF idea for identifying high-quality wine.

Tech Stack & Core Skills

• Deep Learning Framework: PyTorch
• Machine Learning Concepts: Kernel Approximation, Random Fourier Features (RFF), Tabular Classification
• Data Engineering: Feature Standardization, Stratified Splitting, Data Preprocessing
• Model Optimization: Macro F1 Evaluation, Early Stopping, Hyperparameter Comparison
• Monitoring: TensorBoard

Architecture & Methodology

Dataset & Preprocessing The pipeline uses a white wine quality dataset containing physicochemical features (acidity, residual sugar, chlorides, density, pH, sulphates, alcohol) and a quality score.

• Drops duplicate rows and confirms no missing values are present.
• Maps the original quality score into classification labels (0: Low quality <= 5, 1: Medium quality = 6, 2: High quality >= 7).
• Uses stratified train, validation, and test splits.
• Standardizes features with StandardScaler fitted exclusively on the training set.

Model Architecture The model transforms the tabular inputs using fixed random frequencies and phases (cosine and sine features), then trains a fully connected head:

Standardized wine features
  -> fixed Random Fourier Feature mapping
  -> fully connected neural head
  -> class logits
  -> cross-entropy loss

The RFF layer keeps its random projection fixed; only the neural classifier head is trained with gradient descent.

Experimental Variations

Model	Main Change
Baseline RFF	RFF dimension 64, hidden dimension 64, sigma 1.0, learning rate 1e-3.
Variation 1	Same baseline architecture with a smaller learning rate of 5e-4.
Variation 2	Adds L2 regularization with weight decay 1e-4.
Variation 3	Higher-capacity RFF model with 256 RFF features, larger hidden layer, stronger regularization, and early stopping.
Variation 4	Tuned RFF model with 128 RFF features, sigma 0.5, light regularization, and early stopping.

Results & Visuals

Three-Class Classification Results:

Model	Val Accuracy	Test Accuracy	Val Macro F1	Test Macro F1
Baseline RFF	0.5076	0.4857	0.4795	0.4654
Variation 1	0.5479	0.5277	0.5153	0.5079
Variation 2	0.5361	0.5261	0.5034	0.5022
Variation 3	0.5983	0.5361	0.5705	0.5144
Variation 4	0.6067	0.5714	0.5861	0.5553

Variation 4 achieved the best test accuracy and test macro F1 among the three-class RFF experiments.

Binary High-Quality Classification Results:

Task	Val Accuracy	Test Accuracy	Val Macro F1	Test Macro F1
High vs not-high wine quality	0.8084	0.7933	0.6399	0.6228

Repository Structure

rff_wine_quality_classifier.ipynb   Final RFF implementation pipeline
winequality-white.csv               White wine quality dataset
requirements.txt                    Python dependencies

Requirements

If you wish to run this pipeline locally:

• Python dependencies can be installed via pip install -r requirements.txt.
• Open rff_wine_quality_classifier.ipynb to execute the pipeline. The script expects winequality-white.csv to remain in the repository root.
• Optional: Training logs can be viewed locally by running tensorboard --logdir runs.