[Task] Implement Isolation Forest on the Shuttle Dataset

[ankitkumar25-rk]

Pull Request

Related Issue

Closes #18

---

Task Summary

I used Isolation Forest algorithm to find anomalies in the Shuttle dataset. This dataset had 7
classes in it, but I needed only 2 classes (normal and anomaly) for this task, so I changed it
myself. I made class 1 and 4 as normal, and the rest of the classes as anomaly because they had
very less count. After that I trained the Isolation Forest model (this model does not need labels
to train, which is a good thing) and checked how good it is doing by comparing with the real
labels. I also changed the contamination value a few times to see what happens to the results.

---

Dataset

• Mammography
• Shuttle

Dataset Source: I got the dataset from OpenML in .arff file format.

---

Preprocessing

• I checked for missing values using isnull().sum() and there was no missing value anywhere
• I used scipy.io.arff to open the file because normal read_csv was not working on arff files
• The class column was coming like b'1' (in bytes form), so I used .str.decode('utf-8') to
  fix it and then changed it to int
• I made a new label column: class 1 and 4 = normal (0), class 2,3,5,6,7 = anomaly (1). This
  gives around 6.05% anomaly rate which looks realistic
• I took the 9 columns (A1 to A9) as X and label as y (y is only used to check results later, not
  used while training)
• I split the data into train (70%) and test (30%) parts, and used stratify so the anomaly ratio
  stays same in both parts
• I did not do any scaling or encoding because all 9 features were already numbers

---

Model Configuration

Hyperparameter	Value
n_estimators	100
contamination	0.0605
max_samples	256
max_features	1.0 (default, used all features)
random_state	42

---

Evaluation Results

Metric	Value
Precision	0.943
Recall	0.937
F1-score	0.940
ROC-AUC (Optional)	0.996

---

Visualizations

---

Key Observations

• The model worked really good even without using labels while training. It just finds points
  that are easy to separate from others and calls them anomaly
• Contamination value made the biggest difference in the results because it decides where the
  model draws the line between normal and anomaly
  • When contamination was very low (0.01), precision was perfect (1.000) but recall was very bad
    (0.146), because model became too strict
  • When contamination was high (0.2), recall became very good (0.992) but precision became bad
    (0.298), because model started calling too many points as anomaly
  • Best result came at contamination = 0.0605, which is same as the real anomaly rate in data
• I did not change n_estimators and max_samples much because they were already giving good result
  at default values (100 and 256)
• Challenges I faced: Understanding the .arff file and fixing the bytes format class column took some time; deciding
  how to split the 7 classes into 2 classes was confusing at first but looking at the class count
  helped me decide

---

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