twitter-bot-detection

TWITTER Bot Detection Model

Overview

This directory contains a trained Random Forest classifier for detecting bot accounts on Twitter.

Model Version: v2 Training Date: 2025-11-27 12:08:54 Framework: scikit-learn 1.5.2 Algorithm: Random Forest Classifier with GridSearchCV Hyperparameter Tuning

---

📊 Model Performance

Final Metrics (Test Set)

Model Improvement

• Baseline ROC-AUC: 0.9314
• Tuned ROC-AUC: 0.9354
• Improvement: 0.0040 (0.43%)

---

🗂️ Files

---

🎯 Dataset Information

Training Configuration

• Training Samples: 29,951
• Test Samples: 7,487
• Total Samples: 37,438
• Number of Features: 12
• Cross-Validation Folds: 5
• Random State: 42

Class Distribution

Training Set:

• Human (0): 20,028 (66.87%)
• Bot (1): 9,923 (33.13%)

Test Set:

• Human (0): 4,985 (66.58%)
• Bot (1): 2,502 (33.42%)

---

🔧 Features (12)

1. has_custom_cover_image
2. description_length
3. favourites_count
4. followers_count
5. friends_count
6. followers_to_friends_ratio
7. has_location
8. username_digit_count
9. username_length
10. statuses_count
11. is_verified
12. account_age_days

---

🏆 Top 5 Most Important Features

1. followers_count - 0.1895
2. favourites_count - 0.1813
3. friends_count - 0.1494
4. statuses_count - 0.1244
5. account_age_days - 0.1010

---

⚙️ Hyperparameters

Best Parameters (from GridSearchCV)

• class_weight: balanced
• max_depth: 20
• max_features: sqrt
• min_samples_leaf: 1
• min_samples_split: 2
• n_estimators: 300

Parameter Search Space

• n_estimators: [100, 200, 300]
• max_depth: [10, 15, 20, None]
• min_samples_split: [2, 5, 10]
• min_samples_leaf: [1, 2, 4]
• max_features: ['sqrt', 'log2']
• bootstrap: [True, False]

Total combinations tested: 540

---

📈 Cross-Validation Results

Mean Scores (5-Fold Stratified CV)

• Accuracy: 0.8750 (±0.0053)
• Precision: 0.8658 (±0.0089)
• Recall: 0.7368 (±0.0113)
• F1-Score: 0.7961 (±0.0092)
• ROC-AUC: 0.9325 (±0.0037)

---

🖼️ Visualizations

All visualizations are saved in the images/ directory:

1. 01_class_distribution.png - Training/Test set class distribution
2. 02_feature_correlation.png - Feature correlation with target variable
3. 03_correlation_matrix.png - Feature correlation heatmap
4. 04_baseline_confusion_matrix.png - Baseline model confusion matrix
5. 05_baseline_roc_curve.png - Baseline ROC curve
6. 06_baseline_precision_recall.png - Baseline Precision-Recall curve
7. 07_baseline_feature_importance.png - Baseline feature importance
8. 08_cross_validation.png - Cross-validation score distribution
9. 09_tuned_confusion_matrix.png - Tuned model confusion matrix
10. 10_tuned_roc_curve.png - Tuned ROC curve
11. 11_tuned_precision_recall.png - Tuned Precision-Recall curve
12. 12_tuned_feature_importance.png - Tuned feature importance
13. 13_model_comparison.png - Baseline vs Tuned comparison

---

🚀 Usage Example

import joblib
import pandas as pd
import numpy as np

# Load model and scaler
model = joblib.load('twitter_bot_detection_v2.pkl')
scaler = joblib.load('twitter_scaler_v2.pkl')

# Prepare your data (example)
data = {
    'has_custom_cover_image': 0.5,
    'description_length': 0.5,
    'favourites_count': 0.5,
    'followers_count': 0.5,
    'friends_count': 0.5,
    'followers_to_friends_ratio': 0.5,
    'has_location': 0.5,
    'username_digit_count': 0.5,
    'username_length': 0.5,
    'statuses_count': 0.5,
    'is_verified': 0.5,
    'account_age_days': 0.5,
}

# Create DataFrame
df = pd.DataFrame([data])

# Scale features
df_scaled = scaler.transform(df)

# Predict
prediction = model.predict(df_scaled)[0]
probability = model.predict_proba(df_scaled)[0]

print(f"Prediction: {'Bot' if prediction == 1 else 'Human'}")
print(f"Bot Probability: {probability[1]:.4f}")
print(f"Human Probability: {probability[0]:.4f}")

---

📋 Confusion Matrix Breakdown

Tuned Model (Test Set)

                Predicted
              Human    Bot
Actual Human     4676     309
       Bot        611    1891

• True Negatives (TN): 4,676 (Correctly identified humans)
• False Positives (FP): 309 (Humans incorrectly classified as bots)
• False Negatives (FN): 611 (Bots incorrectly classified as humans)
• True Positives (TP): 1,891 (Correctly identified bots)

---

🔍 Model Interpretation

Strengths

• High ROC-AUC score (0.9354) indicates excellent discrimination capability
• Balanced precision and recall for both classes
• Robust cross-validation performance

Key Insights

1. Top features drive bot classification effectively
2. GridSearchCV improved performance over baseline by 0.43%
3. Model generalizes well on unseen test data

---

📝 Notes

• Feature Scaling: All features are scaled using MinMaxScaler to [0, 1] range
• Missing Values: Filled with 0 during preprocessing
• Class Balance: Imbalanced dataset
• Model Type: Ensemble method resistant to overfitting

---

🔄 Model Updates

To retrain the model:

1. Place new training data in ../data/train_twitter.csv
2. Run the training notebook: 5_enhanced_training.ipynb
3. Update this README with new metrics

---

📧 Contact & Support

For questions or issues regarding this model, please refer to the main project documentation.

---

Generated: 2025-11-27 12:08:54 Notebook: 5_enhanced_training.ipynb Platform: Twitter