Telecom QoE Prediction

Business Context

Quality of Experience (QoE) measured as MOS (1-5) directly predicts customer satisfaction and churn risk. Proactively predicting QoE from network KPIs enables real-time service quality management.

Problem Framing

Regression using LightGBM.

• Target: mos_score
• Primary Metric: RMSE
• Challenges:
  • Non-linear MOS curve (diminishing returns at high quality)
  • App-specific sensitivity (gaming vs browsing)
  • Device capability effects on perceived quality

Data Engineering

Session-level data (10K sessions) combining:

• Network KPIs -- SINR, throughput, latency, packet loss
• Device class -- high-end, mid-range, low-end capability tiers
• App type -- gaming, video streaming, browsing, VoIP

Domain physics: MOS computed using an ITU-T inspired formula with app-specific adjustments. Gaming sessions are highly sensitive to latency, video to throughput and packet loss, browsing to page load time proxies.

Methodology

• LightGBM regression with early stopping on validation RMSE
• Feature groups:
  • Network quality index (composite SINR + throughput + latency score)
  • Service degradation indicators (packet loss rate, jitter)
  • Bandwidth utilization (throughput vs expected capacity)
  • App sensitivity score (app-specific quality weight)
• Hyperparameter tuning with cross-validation

Key Findings

• RMSE: 0.45, R²: 0.58 on held-out test set (MOS scale 1-5)
• Top predictors: service_degradation, throughput_mbps, and latency_ms dominate SHAP importance
• Gaming sessions show highest RMSE (0.47) due to latency sensitivity; browsing and VoIP are most predictable
• Unobserved factors (user expectations, content quality, environmental context) limit R² -- realistic for QoE prediction where subjective perception adds irreducible noise

Quick Start

# Clone the repository
git clone https://github.com/adityonugrohoid/telecom-ml-portfolio.git
cd telecom-ml-portfolio/04-qoe-prediction

# Install dependencies
uv sync

# Generate synthetic data
uv run python -m qoe_prediction.data_generator

# Run the notebook
uv run jupyter lab notebooks/

Project Structure

04-qoe-prediction/
├── README.md
├── pyproject.toml
├── notebooks/
│   └── 04_qoe_prediction.ipynb
├── src/
│   └── qoe_prediction/
│       ├── __init__.py
│       ├── data_generator.py
│       ├── features.py
│       ├── model.py
│       └── evaluate.py
├── data/
│   └── .gitkeep
├── models/
│   └── .gitkeep
└── tests/
    └── .gitkeep

Related Projects

#	Project	Description
1	Churn Prediction	Binary classification to predict customer churn
2	Root Cause Analysis	Multi-class classification for network alarm RCA
3	Anomaly Detection	Unsupervised detection of network anomalies
4	QoE Prediction (this repo)	Regression to predict quality of experience
5	Capacity Forecasting	Time-series forecasting for network capacity planning
6	Network Optimization	Optimization of network resource allocation

License

This project is licensed under the MIT License. See LICENSE for details.

Author

Adityo Nugroho