HabibKiz/4g-congestion-detection-prediction

Hybrid ML pipeline for 4G network congestion detection and forecasting. Uses HDBSCAN clustering to segment cell behaviour and XGBoost to predict KPI trajectories 1-7 days ahead. Built on live Orange Mali RAN data. MSc thesis, University of Middlesex Dubai / Sofrecom (Orange Group). # Intelligent Network Monitoring: Clustering-based Detection and Predictive Modelling of Congestion in 4G Networks **MSc Thesis · University of Middlesex Dubai · 2025** **Author:** Habib Kizamou · Internship at Sofrecom (Orange Group), Rabat --- ## Overview Telecom networks generate massive streams of KPI data — but turning that data into early warnings operators can actually act on is the real challenge. This project tackles exactly that, using a **hybrid ML pipeline** combining unsupervised clustering and supervised forecasting on live 4G radio access network data from **Orange Mali**. The pipeline moves from raw operational exports to two operational outputs: 1. **Clustering** — segment network cells by behaviour to surface congestion-risk cohorts 2. **Forecasting** — predict KPI trajectories 1–7 days ahead to enable proactive intervention --- ## Results at a Glance | Task | Best Model | Key Metric | |------|-----------|------------| | Cell segmentation (11 KPIs, 1M+ obs) | HDBSCAN | Silhouette: **0.69–0.70** | | Short-term forecasting (1–7 day) | XGBoost | **30–60% lower MAE** vs Prophet | | Synthetic Busy Hour generation | XGBoost Tweedie | 1 month → **6 months** coverage | --- ## Dataset Data was extracted from **Orange Mali's 4G Radio Access Network** via Sofrecom's PRS interface — live operational KPI exports across four temporal granularities: | Granularity | Rows | |------------|------| | Weekly | ~637K+ (combined) | | Daily | | | Hourly | | | Busy Hour (BH) | | **11 KPIs tracked:** accessibility, retainability, mobility, traffic volume, throughput, and user experience metrics. > ⚠️ **Note:** Raw data is proprietary to Orange Group / Sofrecom and is not included in this repository. Notebooks use anonymised or synthetic data where applicable. --- ## Repository Structure ``` 4g-congestion-detection-prediction/ │ ├── notebooks/ │ ├── 01_data_preparation/ # Cleaning, standardisation, synthetic BH creation │ ├── 02_eda/ …