Applied Data Science Pipeline Power Systems → ML

Inspired by Sayantam Sarkar’s transition from EPC/IIT power engineering to OTT subscriber analytics
State: Active Predictor
The Industrial Engineering Bridge: From Grid Telemetry to OTT Churn Features.

Engineers already understand multi-channel sensor logs, voltage drops, and duty cycles. The mathematical structures directly map to streaming analytics.

1. Customer Telemetry & Features ID: SUB-88492
Avg Watch Hours / Week 4.2 hrs
Power analogy: Grid baseline demand (low hours indicates imminent disconnect)
Days Since Last Login 14 days
Power analogy: Telemetry heartbeat loss duration
Support Tickets (Last 30 Days) 3 tickets
Power analogy: High alarm spike / anomalous transient spikes
Subscription Tenure 8 months
Power analogy: Equipment service age (mid-lifecycle cliff)
Model Selection XGBoost Active
Logistic Reg
Random Forest
XGBoost
Decision Boundary Threshold 0.50
Tuning tradeoff: Recall (catch all churners) vs Precision (minimize wasted discounts)
2. Live Model Inference & Evaluation High Risk
Predicted Churn Probability: 0.78
0.0 (Safe) 0.3 (Watch) 0.6 (Critical) 1.0 (Lost)
Model Accuracy
0.914
ROC - AUC
0.942
Recall Rate
0.892
⚡
Mitigation Protocol
Trigger automated retention discount & personalized content recommendation
Classifier Characteristic Curve (Sensivity vs 1-Specificity)
Cohort Confusion Matrix (N=1,000 Validation Test Set) Threshold = 0.50
Ground Truth Predicted Class
Predicted Active Predicted Churn
Actual Active 670
True Negative		 40
False Positive
Actual Churned 31
False Negative		 259
True Positive
3. Multi-Algorithm Head-to-Head Comparison Benchmarked on 10k Historical OTT Records
Model Name Test Accuracy ROC - AUC Precision Recall Inference Latency Engineering Suitability
Logistic Regression 0.812 0.835 0.745 0.720 0.4 ms High interpretability; linear decision plane. Baseline benchmark.
Random Forest 0.884 0.912 0.835 0.840 8.2 ms Ensemble of bagging trees; handles nonlinear telemetry well.
XGBoost (Optimal) 0.914 0.942 0.881 0.892 3.1 ms Gradient boosted trees; superior generalization across imbalanced churn sets.