Telecom QoE Analytics

Overview

Telecom QoE Analytics is a comprehensive Data Science Practice project utilizing synthetic telecom-digital-twin dataset. It demonstrates end-to-end analytics capability from raw data profiling and rigorous statistical testing to advanced machine learning modeling and strategic troubleshooting, focused on improving Quality of Experience (QoE) in telecommunications networks.

The project implements a six-phase analytics pipeline: (1) Data Profiling & EDA, (2) Statistical Analysis & Causal Inference, (3) ML Regression for QoE Prediction, (4) ML Classification for Degradation Prediction, (5) Unsupervised Learning & Anomaly Detection, (6) Executive Summary & Strategic Insights. All phases prioritize interpretability and actionability over theoretical complexity.

Key findings include: Cell Congestion has massive effect size (Cohen's d = -2.12) on QoE, far outweighing other metrics. XGBoost achieved strong R² performance (0.7247) for QoE prediction. LightGBM achieved high ROC-AUC (0.9645) for degradation classification with excellent recall (0.92). Anomalies cluster around 5 PM busy hour, suggesting peak load correlation.

Key Metrics & Results

Problem Statement

Telecom operators need to understand drivers of Quality of Experience (QoE) degradation to prioritize network investments. Traditional analytics provide correlations but lack causal inference. ML models need interpretability for field engineers. Anomaly detection must balance recall vs precision for SLA compliance.

Business Context

Network operations require actionable insights, not just model predictions. Field engineers need to understand why cells are degraded (feature importance). False negatives (missing outages) are more costly than false positives (false alarms). Strategic recommendations must translate technical findings to business value.

Technical Challenges

• Moving beyond correlation to causal inference
• Ensuring ML model interpretability for business stakeholders
• Handling class imbalance in degradation prediction
• Defining 'normal' in highly dynamic networks for anomaly detection
• Translating technical findings to strategic recommendations

Solution Architecture

A six-phase structured pipeline: (1) Data Profiling with schema validation and QoE distribution analysis, (2) Statistical Analysis with ANOVA and effect size (Cohen's d), (3) ML Regression using XGBoost with Optuna tuning, (4) ML Classification using LightGBM with class imbalance handling, (5) Unsupervised Learning with STL decomposition and Isolation Forest, (6) Executive Summary translating findings to strategic recommendations.

System Components

Statistical Rigor Phase

Hypothesis testing (ANOVA) confirms QoE differences between segments. Effect size analysis (Cohen's d) quantifies impact magnitude. Identified congestion as primary driver (d=-2.12).

ML Regression Pipeline

XGBoost Regressor tuned with Optuna. Achieved R²=0.7247, MAE=0.3672, RMSE=0.4560 on test set. Feature importance identifies latency and congestion as top predictors.

ML Classification Pipeline

LightGBM Classifier handling class imbalance. Achieved ROC-AUC=0.9645 with precision=0.46 and recall=0.92 for minority 'Low QoE' class. Excellent recall ensures proactive intervention. Serves as CEM dashboard engine.

Anomaly Detection System

STL Decomposition for trend/seasonality removal, followed by Isolation Forest. Successfully isolated anomalies (~5% of data) clustering around 5 PM busy hour.

SHAP Interpretability

Game-theoretic feature attribution proving congestion (not just signal strength) is primary QoE driver. Provides explainability for business stakeholders.

Technology Stack Rationale

XGBoost/LightGBM chosen over deep learning for tabular data and superior interpretability. SHAP provides consistent feature attribution vs biased gain metrics. Isolation Forest handles multivariate anomalies. STL decomposition removes temporal patterns. Optuna enables automated hyperparameter tuning.

Implementation Highlights

Key Features

• Six-Phase Analytics Pipeline: Structured approach from EDA to strategic insights with clear progression
• Statistical Rigor: Hypothesis testing and effect size analysis (Cohen's d) for causal inference
• SHAP Interpretability: Game-theoretic feature attribution proving congestion is primary QoE driver
• High-Performance Models: Strong performance with class imbalance handling on synthetic dataset
• Business Translation: Executive summary converts technical findings to strategic recommendations

Challenges & Solutions

Results & Impact

Identified congestion as primary QoE driver (effect size d=-2.12). Achieved strong ML performance on synthetic dataset: XGBoost R²=0.7247, LightGBM ROC-AUC=0.9645. Generated strategic recommendations: prioritize backhaul expansion, optimize latency, deploy proactive alerts. Models demonstrate capability for CEM dashboard deployment.

Lessons Learned

What Worked Well

• Tree-based models (XGBoost/LightGBM) excelled on tabular telecom data
• SHAP provided superior interpretability vs standard gain metrics
• Effect size analysis (Cohen's d) identified true drivers vs correlations
• STL decomposition effectively removed temporal patterns for anomaly detection
• Structured six-phase pipeline demonstrated end-to-end thinking

What I'd Do Differently

• Would add temporal cross-validation for time-series leakage prevention
• Should implement online learning for model adaptation over time
• Could add ensemble methods for improved robustness
• Would add A/B testing framework for recommendation validation

Future Enhancements

• Add temporal cross-validation for time-series models
• Implement online learning for continuous model improvement
• Build CEM dashboard integration for real-time QoE monitoring
• Add A/B testing framework for strategic recommendation validation

Related Projects