Firefly

Overview

• This project uses the Firefly Algorithm to detect anomalies in the UCI Red Wine Quality dataset.
• Anomalies are wines with extremely low quality or extremely high quality.
  • We drew the line at quality <= 3 and quality >= 8 to be considered an anomoly.
• The dataset has 1599 samples with 12 physicochemical features such as acidity, alcohol, density, and pH. Only about 1.75 percent of the samples are anomalies, making the data highly imbalanced.

Method

• Anomalies are detected using a weighted Mahalanobis distance that emphasizes the most relevant features.
• The Firefly Algorithm optimizes feature weights and a percentile threshold to maximize cross-validated F1 score, average precision, and ROC-AUC while penalizing large weights for interpretability.
• The algorithm simulates fireflies moving toward brighter solutions, using 80 fireflies over 60 generations with moderate randomness decay.

Results

• The algorithm highlighted volatile acidity, alcohol, and density as the most important features.
• Example metrics include precision 0.68, recall 0.71, F1 score 0.69, ROC-AUC 0.93, and average precision 0.79.
• Visualizations include FA convergence, feature weight distribution, PCA projection highlighting anomalies, score distribution with threshold, and precision-recall curves.

Full Report

For detailed explanations, mathematical justification, and all visualizations, see the FA-report.pdf and our Jupyter notebook file located in this repository.