🧠 Predicting Blockchain Transaction Types using ML Models

A machine learning project to classify blockchain transactions into sales, purchases, transfers, scams, and phishing attempts.

---

📋 Project Overview

This project explores the application of various machine learning models to predict transaction types within blockchain data. The dataset was sourced from Kaggle and contained ~78,600 records. The primary goal is to improve fraud detection in blockchain transactions using effective classification techniques.

🚀 Features

✅ Implemented multiple ML models for classification:
• Random Forest
• K-Nearest Neighbors (KNN)
• Gaussian Naive Bayes
• Decision Tree
• AdaBoost
✅ Applied effective preprocessing techniques like: • Oversampling, Undersampling, Stratified Learning, and SMOTE
• Information Gain for feature selection
• Min-Max Normalization for feature scaling
✅ Evaluated model performance using metrics like accuracy, precision, recall, and F1-score.

🗄️ Dataset

• Source: Kaggle Blockchain Dataset
• Size: ~78,600 records
• Attributes: Various transaction details, including timestamps, regions, and behavioral patterns.

Data Imbalance Challenge: The dataset had a skewed distribution with legitimate transactions heavily outnumbering fraudulent ones. Techniques such as oversampling, undersampling, and SMOTE were implemented to mitigate this issue.

🔍 Data Preprocessing

• Dropped Irrelevant Features: Removed attributes like timestamp, sending_address, and receiving_address.
• Handling Missing Values: Used median imputation for numerical values.
• Categorical Encoding: Applied Label Encoding to variables such as location_region, purchase_pattern, age_group, and anomaly.
• Feature Scaling: Used Min-Max Normalization to ensure consistent feature scaling.

⚙️ Models & Hyperparameter Tuning

Model	Key Hyperparameters	Best Accuracy
AdaBoost	n_estimators, learning_rate	72.9%
Gaussian Naive Bayes	var_smoothing	69.2%
KNN	n_neighbors, weights	97.8%
Random Forest	n_estimators, min_samples_split	97.9%

✅ Best-performing Model: Random Forest Classifier

📊 Results & Analysis

• Random Forest achieved the highest accuracy of 97.9% and demonstrated consistent performance across various configurations.
• KNN closely followed with an accuracy of 97.8%, excelling particularly in 'purchase' and 'sale' transactions.
• The AdaBoost and Gaussian Naive Bayes models struggled with certain transaction types, such as 'phishing' and 'scam'.

🔮 Future Improvements

📈 Exploring additional ensemble models like XGBoost and Gradient Boosting.
🧹 Investigating advanced preprocessing techniques for better feature extraction.
🧠 Improving model performance for minority class predictions through enhanced sampling techniques.

🤝 Contributors

• Omnia Osama Ahmed
• Sara Imad Hamdan
• Nour Bashar Soukieh

📄 Course Work

This project was developed for the course Artificial Intelligence (CSC406) In Abu Dhabi Uniersity.

---

💻 Built with ❤️ for secure digital transactions in the Metaverse!