Add Isolation Forest notebook

[khyati-1711]

Pull Request

Related Issue

Closes #14

---

Task Summary

Implemented an anomaly detection pipeline using Isolation Forest on the Mammography dataset to identify (anomalous) cases .

• Loaded the Mammography dataset via fetch_openml, cleaned it (nulls, duplicates), and mapped the target to a binary scheme (0 = normal, 1 = anomaly/malignant).
• Split the data into train/test sets (stratified) and scaled features with StandardScaler.
• Trained an IsolationForest model on the scaled training data and used it to score/predict anomalies on the full deduplicated dataset.
• Evaluated performance using precision, recall, F1, ROC-AUC, a classification report, and a confusion matrix.
• Approach followed: standard unsupervised outlier-detection workflow (EDA → cleaning → scaling → Isolation Forest fit → scoring → threshold-based prediction → evaluation against ground-truth labels).

---

Dataset

• Mammography
• Shuttle

Dataset Source: sklearn.datasets.fetch_openml(name="mammography", version=1, as_frame=True)

---

Preprocessing

• Missing values: Checked with isnull().sum() (none found), and dropna() applied .
• Duplicates: 3,334 duplicate rows identified and dropped (dataset went from 11,183 → 7,849 rows).
• Feature scaling: StandardScaler fit on training features and applied to both train and test sets.
• Encoding: Target labels mapped from {-1, 1} to {0, 1} (0 = normal, 1 = anomaly).
• Feature selection: None performed — all 6 original attributes (attr1–attr6) retained.
• Train/test split: 80/20 split, stratified on the target class, random_state=42.

---

Model Configuration

Hyperparameter	Value
n_estimators	300
contamination	0.03
max_samples	auto
max_features	1.0
random_state	42

---

Evaluation Results

Metric	Value
Precision	0.189
Recall	0.256
F1-score	0.217
ROC-AUC	0.821

Confusion Matrix:

Classification Report (per class):

Class	Precision	Recall	F1-score	Support
0 (Normal)	0.98	0.96	0.97	7595
1 (Anomaly)	0.20	0.27	0.23	254
Accuracy			0.94	7849

---

Visualizations

Attach at least 2 plots from your analysis.

1. Class Distribution bar plot

2. Anomaly Score Scatter Plot

3. Confusion Matrix heatmap

---

Key Observations

• What worked well: The pipeline runs end-to-end cleanly — data loading, cleaning, scaling, model fitting, and evaluation all execute without errors. ROC-AUC of ~0.82 shows the model has reasonable ranking ability between normal and anomalous points despite no label supervision during training.
• Biggest impact hyperparameter: contamination (set to 0.03) directly controls the anomaly threshold and therefore has an effect on precision/recall trade-off — since the true anomaly rate (~3.2%) is close to this value.
• Challenges faced: Significant duplicate rows (3,334 of 11,183) needed removal before training to avoid biasing the fitted distribution. Care was also needed to keep X/y alignment consistent after deduplication and label remapping across -1/1 and 0/1 conventions.

---

Checklist

• Code runs successfully
• Notebook (.ipynb) included
• Code is well-commented (minimal comments — consider adding more before merge)
• README/documentation updated
• At least 2 plots included
• PR is linked to the corresponding issue