• Overview
• Research Question & Hypothesis
• Dataset
• ML Pipeline
• Project Phases
• Results
• Key Findings
• Tech Stack
• How to Run
• Repository Structure
• Author

This project builds a complete end-to-end machine learning pipeline to predict downlink throughput (Quality of Service) in a 5G Radio Access Network (RAN).

Using real-world drive-test measurements from the Berlin V2X vehicular scenario, the project progresses through five structured phases — raw data collection, cleaning, exploratory analysis, feature engineering, and model training — culminating in a high-accuracy ensemble model.

Can we use 5G radio, mobility, and context features to predict the downlink throughput of a vehicular UE under Carrier Aggregation (CA)?

H1 (Testable): If radio link quality and allocated resources improve — indicated by higher PCell_SNR, RSRP, and Downlink_Num_RBs — then the downlink data rate (target) will increase significantly.

Result: Hypothesis confirmed. Radio signal metrics and resource block allocation are the dominant predictors of 5G throughput.

Raw Drive-Test CSV
 │
 ▼
┌─────────────────────┐
│ Phase 1 │ Data ingestion, shape & type exploration
│ Data Collection │
└────────┬────────────┘
 │
 ▼
┌─────────────────────┐
│ Phase 2 │ Duplicate removal, KNN + domain-driven imputation,
│ Cleaning & Prep │ 3GPP-bounded outlier clipping
└────────┬────────────┘
 │
 ▼
┌─────────────────────┐
│ Phase 3 │ Distributions, correlation heatmaps,
│ EDA │ mutual information, Simpson's paradox, seasonality
└────────┬────────────┘
 │
 ▼
┌─────────────────────┐
│ Phase 4 │ PCA (95% variance), LassoCV feature selection,
│ Feature Eng. │ signal ratios, temporal cyclical encoding
└────────┬────────────┘
 │
 ▼
┌─────────────────────┐
│ Phase 5 │ Ridge, SVR, Decision Tree, XGBoost, LightGBM,
│ Modeling │ Random Forest, Voting Regressor + GridSearchCV
└────────┬────────────┘
 │
 ▼
 R² = 0.960 

1. Radio signal quality drives throughput — SINR, RSRP, and allocated resource blocks (Num_RBs) are the top predictors
2. Weather is a near-zero predictor — confirmed by both mutual information analysis and LassoCV coefficient shrinkage
3. Ensemble beats individuals — Voting Regressor (R²=0.96) outperforms standalone XGBoost (R²=0.935) by combining diverse learners
4. No overfitting — learning curves show train/validation convergence across all models
5. Simpson's Paradox observed — area-level aggregation can reverse individual feature-target trends, emphasizing the need for stratified analysis

Run environment: Google Colab with T4 GPU

1. Open 602.p5.SriRammSekarSasirekha.ipynb in Google Colab
2. Mount Google Drive
3. Upload Train.csv, Test.csv, and VariableDefinitions.csv to your Drive
4. Update the BASE path at the top of the notebook:

BASE = "/content/drive/MyDrive/YOUR_FOLDER_NAME"

5. Set runtime to T4 GPU → Runtime → Change runtime type → GPU
6. Run all cells sequentially (Runtime → Run all)

# Clone the repo
git clone https://github.com/SriRammSS/5GQoS-Analysis.git
cd 5GQoS-Analysis

# Install dependencies
pip install -r requirements.txt

# Launch Jupyter
jupyter notebook 602.p5.SriRammSekarSasirekha.ipynb

Note: The notebook mounts Google Drive for data loading. For local runs, update the BASE variable to a local path containing the CSV files.

5GQoS-Analysis/
│
├── 602.p5.SriRammSekarSasirekha.ipynb # Full 5-phase project notebook
├── requirements.txt # Python dependencies
├── LICENSE # MIT License
└── README.md # Project documentation

Sri Ramm Sekar Sasirekha UID: 121949820 | Course: 602 — Principles of Data Science