Supervised & Unsupervised Learning — Class Imbalance and SMOTE

---

Overview

A focused study of class imbalance in supervised learning — one of the most common and consequential problems in applied ML. When one class is rare (fraud = 0.17%, disease = 1%), standard classifiers optimize toward the majority class and fail on the cases that matter most.

This sprint covers two datasets:

1. Customer churn prediction (Churn.csv) — telecom subscriber churn with moderate imbalance
2. Credit card fraud detection — severe imbalance (1:100 class ratio), addressed with SMOTE

---

The Class Imbalance Problem

Naive classifier that predicts "no fraud" on all examples:
  → Accuracy: 99.83%  ← looks great
  → Recall:    0.0%   ← catches zero fraud cases
  → AUC-ROC:   0.5    ← equivalent to random guessing

Why accuracy is the wrong metric for imbalanced data.

---

SMOTE — Synthetic Minority Oversampling Technique

Rather than duplicating existing minority samples (which adds no new information), SMOTE synthesizes new samples by interpolating between real minority-class examples in feature space:

Algorithm:
1. Select a random minority-class example x
2. Find k nearest neighbors of x in feature space (typically k=5)
3. Randomly choose one neighbor x_neighbor
4. Generate synthetic example:
       x_new = x + λ × (x_neighbor - x)
       where λ ~ Uniform(0, 1)

Result: new examples along line segments between real minority samples

from imblearn.over_sampling import SMOTE

sm = SMOTE(random_state=123)
X_train_res, y_train_res = sm.fit_resample(X_train, y_train)

# Before SMOTE:  {0: 199020, 1: 345}     ← 577:1 ratio
# After SMOTE:   {0: 199020, 1: 199020}  ← balanced

---

Full Modeling Pipeline

1. Data Loading
   └── Churn.csv / creditcard.csv → pandas

2. Feature Engineering
   ├── StandardScaler on 'Amount' column
   └── Drop 'Time' (non-informative)

3. Train/Test Split
   └── 70% train / 30% test (stratified)

4. SMOTE Oversampling
   └── Applied to TRAINING set only (never to test set)

5. Model Training
   └── Logistic Regression (C=4 regularization)

6. Evaluation
   ├── ROC-AUC score
   ├── Confusion matrix (seaborn heatmap)
   ├── Precision-Recall at multiple thresholds
   │       threshold in np.arange(0, 1, 0.02)
   └── ROC curve (FPR vs. TPR)

---

Key Evaluation Metrics

Metric	Formula	When to use
Accuracy	TP+TN / total	Only for balanced classes
Precision	TP / (TP+FP)	Minimize false alarms
Recall (Sensitivity)	TP / (TP+FN)	Minimize missed positives
ROC-AUC	Area under ROC curve	Overall discrimination ability
F1 Score	2 × (P×R)/(P+R)	Balance precision and recall

---

Files

Supervised-Unsupervised-Learning/
├── Churn.csv                   # Telecom customer churn data (693KB)
└── SMOTE in Supervised Learning  # SMOTE theory + full credit card fraud pipeline

---

Computational Biology Connection

Class imbalance is ubiquitous in biomedical data and the same techniques apply directly:

Supervised/SMOTE concept	Biomedical equivalent
Fraud vs. non-fraud (1:577)	Pathogenic variant vs. benign (1:1000+)
SMOTE synthetic oversampling	Data augmentation in rare disease cohorts
Logistic regression + SMOTE	Variant pathogenicity classifiers (ClinVar, CADD)
Precision-recall threshold sweep	Clinical sensitivity / specificity tradeoff
ROC-AUC	Standard metric in diagnostic model validation
Confusion matrix	Clinical trial outcome categorization (TP/FP/TN/FN)

In GWAS and genomic ML, rare variants and rare disease phenotypes create extreme class imbalance. SMOTE and cost-sensitive learning are active research areas in:

• Rare variant association testing
• Cell type classification in scRNA-seq (rare cell types)
• Pathogenic variant prediction (ClinVar positive:negative ratio ~1:50)
• Clinical outcome prediction in small-N rare disease trials