📄 Related Publication

---

This repository contains the experimental code used in the following article:

Lee, K., Lee, J., Kong, I. et al. (2025)
On the use of supervised anomaly detection algorithms for extremely imbalanced data.
Journal of the Korean Statistical Society (2025)
DOI: https://doi.org/10.1007/s42952-025-00347-x

🎯 Experiment Objective

---

To compare and analyze anomaly detection methods with imbalanced binary classification methods under extremely imbalanced settings.

🧭 Pipeline Overview

---

Data loading → train/validation/test split → (optional) adjustment of minority class size → over/under-sampling → MinMax scaling → model training and score generation → metric storage.

▶️ Example Command

---

python experiment.py -d 35 -m 0 -s 0 -l 3 -mi 3 -se 0 -g 0 --epochs 50 --lr 0.001 --batch_size 128 --thresholdq 0.95

⚙️ Key Parameter Descriptions

---

• -d Dataset ID

• -m Model (0: mlp, 1: lr, 9: rf, 10: svm, 4–7: Anomaly methods)

• -s Sampling method (none/smote/adasyn, etc.)

• -l Loss function (focal, etc.)

• -mi Number of minority (anomaly) samples

• -se Seed preset

• -g GPU

• --epochs / --lr / --batch_size Training configuration

• --thresholdq Quantile threshold (q)

🏷️ Label Convention

---

Anomaly: y=1, Normal: y=0. Scores are assumed to follow “higher = more anomalous.”