Breast Cancer Wisconsin Classification

This project implements a machine learning pipeline to classify tumors as Benign or Malignant using the Breast Cancer Wisconsin dataset.
It covers data cleaning, exploratory analysis, class balancing (SMOTE), and evaluation of models in both Scikit-learn and PyTorch, with a focus on maximizing recall for medical diagnosis.

---

Project Overview

• Goal: Predict tumor type (0 = Benign, 1 = Malignant) from cytological features.
• Dataset: Breast Cancer Wisconsin dataset — 449 samples, 9 integer-valued features.
• Challenge: Class imbalance (~63% benign, ~37% malignant) and minimizing false negatives.
• Approach: Data preprocessing, SMOTE balancing, model comparison, PyTorch implementation with regularization.

---

Steps & Methodology

1. Data Preprocessing

• Missing Values: Imputed using IterativeImputer to preserve multivariate relationships.
• Class Encoding: 2 → 0 (Benign), 4 → 1 (Malignant).
• Outlier Removal: IQR-based filtering reduced dataset from 449 → 317 samples.
• Feature Scaling: All features scaled to [0,1] using MinMaxScaler.
• Train-Test Split: 80% train / 20% test with stratification.

2. Exploratory Data Analysis (EDA)

• Feature Distributions: Malignant tumors generally have higher values in Clump_thickness, Uniformity_of_cell_size, etc.
• Correlation: Strong correlations observed (Uniformity_of_cell_size ↔ Uniformity_of_cell_shape: r = 0.88).
• Class Imbalance: Resolved using SMOTE on training data.

3. Models Trained

• Scikit-learn Models:
  • Perceptron (baseline)
  • MLP (default & tuned via GridSearchCV)
• PyTorch Models:
  • Custom MLP architecture
  • Tuned hyperparameters (hidden layers, activation, learning rate)
  • Regularization experiments (Dropout, L2 weight decay)

---

Final Results

Performance Summary (Test Set)

While all models achieved AUC > 0.97, the most important factor in a medical setting is recall for the malignant class.

• Best Scikit-learn Model: Tuned MLP — 83.3% recall, AUC 0.9771.
• Best PyTorch Model (No Reg): Recall 96%, strong precision and F1.
• PyTorch + Regularization:
  • Dropout (p=0.5) → 100% recall (zero false negatives), Accuracy 95.3%, F1-score 0.9412, AUC 0.9844.
  • L2 regularization also boosted performance, but Dropout achieved the best trade-off.

Champion Model:

PyTorch MLP — single hidden layer (8 neurons), LeakyReLU activation, learning rate 0.001, Dropout rate 0.5.

Why It Wins:

• Perfect Recall (100%) — no malignant tumors missed.
• Highest Accuracy among perfect-recall models (95.3%).
• Best F1-Score (0.9412) — ideal precision/recall balance.
• Top AUC (0.9844) — excellent class separation ability.