Model Card: Gradient Boosting Classifier

Purpose: Binary classification of breast cancer tumors (malignant vs. benign) using the Wisconsin Diagnostic Breast Cancer dataset.

Dataset Characteristics

Characteristic	Value
Total Features	30
Training Samples	398 (Class 0: 148, Class 1: 250)
Test Samples	86 (Class 0: 32, Class 1: 54)
Class Balance (Train)	62.8% positive class

Feature Importance Analysis

Top 10 most influential features in the model's decision-making:

Rank	Feature	Importance Score
1	`worst radius`	0.3663
2	`worst concave points`	0.3431
3	`worst area`	0.0950
4	`worst perimeter`	0.0936
5	`concavity error`	0.0150
6	`worst texture`	0.0134
7	`mean concave points`	0.0117
8	`mean texture`	0.0108
9	`worst concavity`	0.0103
10	`texture error`	0.0083

Interpretation: The model relies most heavily on worst radius, which contributes 36.6% to the overall feature importance. This suggests this measurement is particularly discriminative for distinguishing malignant from benign tumors.

Model Behavior Analysis

Confusion Matrix (Test Set)

	Predicted: Benign	Predicted: Malignant
Actual: Benign	27	5
Actual: Malignant	3	51

Error Analysis:
• False Positives (benign → malignant): 5 cases
• False Negatives (malignant → benign): 3 cases

Note: In medical contexts, false negatives are typically more concerning as they represent missed cancer diagnoses.

Model Architecture

Gradient Boosting Configuration:
• Ensemble Size: 200 trees
• Learning Rate: 0.05
• Tree Depth: 2 (shallow trees for regularization)
• Subsampling: 1.0 (uses all training data per iteration)

Intended Use Cases

Primary: Educational demonstration of ML experiment tracking
Research: Baseline model for breast cancer classification benchmarks
Prototyping: Template for clinical ML workflows

Limitations & Caveats

Important Limitations:

No Calibration: Probability outputs may not be well-calibrated. Consider isotonic or Platt scaling for clinical use.
Fixed Threshold: Uses default 0.5 decision threshold without cost-benefit analysis for medical context.
No Cross-Validation: Single train-test split may not capture full model variance.
Dataset Bias: Wisconsin dataset may not generalize to different populations, imaging equipment, or clinical protocols.
Feature Scale Sensitivity: Model may be sensitive to feature scaling differences in production data.
No Fairness Audit: No analysis of performance across demographic subgroups (age, race, etc.).

Reproducibility

Random Seed: 42
Data Splitting: Stratified splits to preserve class distribution
Validation: Staged predictions used for validation curves (per-iteration metrics)

Environment details (sklearn version, numpy version, Python version) should be captured by your experiment tracking backend or requirements.txt.

References & Context

Dataset: Wisconsin Diagnostic Breast Cancer (WDBC) from UCI ML Repository
Features: Computed from digitized images of fine needle aspirate (FNA) of breast mass
Feature Engineering: Mean, standard error, and "worst" values for 10 real-valued characteristics

For detailed metrics and training curves: Refer to the experiment tracking dashboard. This model card focuses on interpretability, data characteristics, and deployment considerations not captured in raw metrics.