Anomaly Detection Experiment Summary

This notebook documents experiments on Anomaly Detection using three unsupervised models on the iot dataset.

Models Implemented

• Isolation Forest
• Local Outlier Factor (LOF)
• One-Class SVM

🛠️ Workflow Overview

1. Data Loading & Exploration

   • Imported dataset
   • Checked for missing values and performed basic data analysis

2. Preprocessing

   • Scaled features with MinMaxScaler
   • Prepared data for anomaly detection tasks

3. Model Training & Prediction

   • Trained each model separately
   • Obtained predictions: -1 for anomaly, 1 for normal observations

4. Performance Evaluation

   • Used Confusion Matrix and Classification Report for each model
   • Focused on precision, recall, and F1-score for anomalies detection (-1 class)

---

📊 Results Summary

Model	Anomaly Precision	Anomaly Recall	Anomaly F1-Score	Accuracy
Isolation Forest	0.20	1.00	0.34	0.97
Local Outlier Factor (LOF)	0.00	0.00	0.00	0.99
One-Class SVM	0.08	1.00	0.14	0.90

---

📌 Observations & Insights

• Isolation Forest showed the best compromise between anomaly detection and overall accuracy. It captured all anomalies but with low precision (many false positives).
• LOF failed to detect anomalies (0 precision and recall) despite achieving high overall accuracy — likely due to extreme imbalance and model sensitivity.
• One-Class SVM detected all anomalies (100% recall) but with even lower precision than Isolation Forest, leading to many false positives and a drop in overall accuracy.

---

✅ Conclusion

• Isolation Forest is the recommended choice for this dataset when maximizing anomaly recall is critical.
• Model selection should balance recall and precision depending on the real-world cost of false positives vs. false negatives.

---