Task: Isolation Forest model on Shuttle dataset

[lalitgirgaonkar]

Pull Request

Related Issue

Closes #15

---

Task Summary

Provide a brief overview of your implementation.

• Implemented Isolation Forest model on the Shuttle dataset from OpenML.
• Loaded the dataset and analysed all the columns. Decided which classes to consider as normal values based on percentage of occurrence. Classes 1 and 4 made up roughly 94% of observations, so there were roughly 6% anomalies.
• Scaled features using RobustScaler.
• Evaluated model performance using precision, recall, F1 scores and ROC-AUC value.
• Plotted Anomaly Distribution and Confusion Matrix

---

Dataset

• Mammography
• Shuttle

Dataset Source: OpenML

---

Preprocessing

Describe any preprocessing performed.

• There were no missing values in the dataset.
• Scaled the data using RobustScaler, to ensure that that extreme anomalies do not affect the values of normal points.
• Mapped classes {1,4} to 1 (normal) and {2,3,5,6,7} to -1 (anomaly)

---

Model Configuration

List the important hyperparameters used.

Hyperparameter	Value
n_estimators	100
contamination	0.0605
max_samples	256
max_features	1 (default)
random_state	42

---

Evaluation Results

Metric	Value
Precision	0.93559
Recall	0.93506
F1-score	0.93532
ROC-AUC (Optional)	0.995

---

Visualizations

Attach at least 2 plots from your analysis.

Anomaly Distribution

Confusion Matrix

---

Key Observations

• The approximation that Class 1 and 4 are normal values had a positive impact, as the values of Precision and Recall were >90%.
• I tried the model with and without scaling. There was not much difference between the two methods.
• Contamination value had an impact on the output. Too small a value of contamination limited the number of anomalies flagged by the model which increased the number of false negatives.
• Using the exact value of contamination gave the same value of Precision and Recall (around 0.93)

---

Checklist

• Code runs successfully
• Notebook (.ipynb) included
• Code is well-commented
• README/documentation updated
• At least 2 plots included
• PR is linked to the corresponding issue