Comparative Analysis of Deep Learning Models for Web Traffic Time-Series Forecasting

A comprehensive comparison of six deep learning architectures (RNN, LSTM, GRU, and their bidirectional variants) for predicting Wikipedia web traffic using time-series analysis.

📄 Project Paper

You can read the full project paper here.

📊 Research Overview

This study evaluates the performance of different recurrent neural network architectures for web traffic forecasting using real-world Wikipedia page view data. The research demonstrates that Bidirectional LSTM (BiLSTM) achieves superior performance with an R² score of 0.93, making it the most effective model for capturing long-term temporal dependencies in web traffic patterns.

🎯 Key Findings

Model	MAE	R² Score	MSE	MSLE
BiLSTM	0.0163	0.9309	0.0018	0.0007
RNN	0.0122	0.7828	0.0023	0.0008
GRU	0.0402	0.7491	0.0041	0.0025
BiGRU	0.0410	0.7390	0.0043	0.0025
BiRNN	0.0260	0.7020	0.0046	0.0020
LSTM	0.0458	0.3291	0.0097	0.0043

🚀 Features

• Comprehensive Model Comparison: Six different RNN architectures tested under identical conditions
• Real-world Dataset: Wikipedia web traffic data from Kaggle (145K+ articles, July 2015 - Dec 2016)
• Multiple Evaluation Metrics: MAE, MSE, MSLE, and R² score for thorough performance assessment
• Bidirectional Processing: Enhanced temporal pattern recognition through forward and backward sequence processing
• GPU Acceleration: Optimized training using NVIDIA Tesla P100 GPU

🔬 Methodology

Data Preprocessing

• Missing Value Handling: Forward fill (ffill) method
• Normalization: Min-Max scaling to [0,1] range
• Sliding Window: Fixed-length sequences for temporal pattern learning
• Train-Test Split: Chronological 80-20 split

Model Architecture

All models follow a consistent architecture:

• 2 recurrent layers (hidden_size=32, dropout=0.2)
• Fully connected layers (64 units → output)
• ReLU activation with dropout regularization

Training Configuration

• Optimizer: Adam (lr=0.001)
• Loss Function: Mean Squared Error (MSE)
• Epochs: 500
• Hardware: NVIDIA Tesla P100 GPU

📈 Results Analysis

Key Insights

1. BiLSTM Superiority: Achieved the highest R² score (0.9309) and lowest MSE/MSLE
2. Bidirectional Advantage: All bidirectional variants outperformed their unidirectional counterparts
3. RNN Efficiency: Simple RNN showed competitive MAE performance despite architectural simplicity
4. LSTM Underperformance: Unexpected poor performance possibly due to hyperparameter sensitivity

Performance Visualization

The repository includes comprehensive visualizations:

• Model comparison charts
• Prediction vs. actual value plots
• Training loss curves
• Error distribution analysis

🔮 Future Work

• Hyperparameter optimization for individual models
• Attention mechanism integration
• Transformer-based architectures exploration
• Multivariate forecasting with exogenous features
• Real-time prediction system development

📚 Citation

If you use this work in your research, please cite:

@article{geddam2024web,
  title={Comparative Analysis of Deep Learning Models for Web Traffic Time-Series Forecasting},
  author={Geddam, Poorvik Shrinil},
  journal={VIT-AP University},
  year={2025}
}

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Kaggle for providing the Web Traffic Time Series Forecasting dataset.
• PyTorch team for the deep learning framework.

---

⭐ Star this repository if you found it helpful!