closes #12 - Abhishek

[abhishekreddy22]

Pull Request

Related Issue

Closes # 12

---

Task Summary

Provide a brief overview of your implementation.

• What did you implement?
  Implemented the Isolation Forest model for detecting anomalous mammography data

• What approach did you follow?

I extracted the data from a csv file,
selected the features to detect the anomaly
used dropna() to eliminate rows with missing values
used StandardScaler to scale features
split the data into test and train (test=0.1)
trained the isolation forest model
used iso_forest.predict() to get the number of anomaly
used matplotlib to plot the data

---

Dataset

• Mammography
• Shuttle

Dataset Source: OpenML

---

Preprocessing

Handled missing values with pandas' built in function df.dropna()
which drops rows with NaN

Found out using SHAP, that col 4/5/6 were the major contributors to the model and rest of the columns were just noise, omitting col 1/2/3 increased f1_score drastically

Omitted using Standard Scaler as the the data was already scaled

---

Model Configuration

List the important hyperparameters used.

Hyperparameter	Value
n_estimators	100
contamination	0.023
max_samples	256
max_features	3
random_state	42

---

Evaluation Results

Metric	Value
Precision	0.615
Recall	0.473
F1-score	0.535
ROC-AUC (Optional)

---

Visualizations

Attach at least 2 plots from your analysis.

Confusion Matrix

Scatter Plot

---

Key Observations

Briefly summarize:

• What worked well?
  The most important function that helped tune the model better was precision_recall_curve(), which helped me know the exact threshold value, instead of blindly going for predict()

• Which hyperparameter had the biggest impact?
  3 Main hyperparameters Contamination|n_estimators|max_samples had high impacts on f1score, especially Contamination and max_features

• Any interesting findings?\

• Challenges faced (if any)
  Most of the challenge faced was getting familiar with the syntax of different libraries, and researching how to improve the f1_score took a lot of time
  initial f1_score: 0.264
  final f1_score: 0.535

Understanding how to evaluate the model (i.e implementing the precision_score etc) was a big challenge due to different conventions among OpenML|Isolation_Forest|Precision_Score

---

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