<![CDATA[The increasing use of marketplace platforms has raised the risk of transaction fraud. Fraud detection is challenging because fraudulent transactions are far less frequent than legitimate transactions, creating an imbalanced dataset. This study analyzes the capability of Isolation Forest for fraud detection, evaluates the influence of preprocessing and hyperparameter settings, and compares its performance with Local Outlier Factor (LOF). The experiment uses the public Fraudulent E-Commerce Transactions dataset from Kaggle containing 1,048,575 transactions: 996,055 legitimate transactions (94.99%) and 52,520 fraudulent transactions (5.01%). The workflow includes data cleaning, feature selection, feature engineering, One-Hot Encoding, StandardScaler, an 80:20 stratified train-test split, baseline modeling, hyperparameter tuning, threshold tuning, evaluation, and comparison with LOF. Isolation Forest is trained in an unsupervised manner, while the Is Fraudulent label is used only as evaluation ground truth. The final configuration uses contamination 0.05, 250 estimators, max_samples 0.80, threshold 0.45, and random_state 42. The model achieves Accuracy 0.9820, Precision 0.8201, Recall 0.8200, F1-Score 0.8200, ROC-AUC 0.9187, and PR-AUC 0.6458. The confusion matrix contains 197,321 True Negatives, 8,613 True Positives, 1,890 False Positives, and 1,891 False Negatives. Isolation Forest outperforms LOF on most metrics, although LOF provides slightly higher Recall. The detection result is intended as an early indicator for transactions requiring further review.]]>
Copyrights © 2026
