Model Performance

VPS নির্বাচন: V1009 V1003 V1001 V1011

মূল ফলাফল — V1001

CPU RMSE (Online)

5.6780%

After online learning

RAM RMSE (Online)

2.9948%

After online learning

RAM R²

0.798

Coefficient of determination

AUC-ROC

0.843

Attention LSTM anomaly

Baseline তুলনা — CPU Forecast RMSE (raw target, 5-min horizon)

* Negative R² reflects concept drift between train/test periods — see Online Learning page

vs Naive Baseline (CPU)

Naive: 18.02% → Ours: 14.65%

18.7% ↓

vs ARIMA (2,1,2) — CPU

ARIMA: 13.49% → Ours: 14.65%

-8.6%

vs Simple LSTM — CPU

LSTM: 13.78% → Ours: 14.65%

-6.3%

vs ARIMA — RAM Forecast

ARIMA: 2.88% → Ours: 0.81%

71.7% ↓

Multi-Horizon Prediction — XGBoost vs Naive Persistence

✅ সব horizon-এ XGBoost naive-কে হারিয়েছে — longer horizon-এ সুবিধা আরও বেশি

Horizon	XGB RMSE	Naive RMSE	XGB জিতেছে?	Improvement	RAM RMSE	RAM R²	Status
5 মিনিট	14.76%	18.12%	✅ হ্যাঁ	18.5% ভালো	0.81%	0.798	Production
10 মিনিট	14.15%	18.85%	✅ হ্যাঁ	24.9% ভালো	1.06%	0.656	সক্ষম
15 মিনিট	13.77%	18.64%	✅ হ্যাঁ	26.1% ভালো	1.2%	0.561	সক্ষম
30 মিনিট	13.54%	17.37%	✅ হ্যাঁ	22.0% ভালো	1.45%	0.355	Experimental

💡 Note: Negative R² reflects concept drift between train/test periods — model still beats naive baseline on RMSE. See Online Learning page for retraining results.

Decision Distribution — 7915 Windows

✅ Normal 5704 72.1%

👁 Monitor 1769 22.3%

⚠️ Scale Soon 922 11.6%

🚨 Scale Now 241 3.0%

Novel Contributions

✅

Real KVM VPS production data

Synthetic dataset নয় — আসল hosting business থেকে

✅

Hybrid PyTorch LSTM + XGBoost

Anomaly detection + forecasting একসাথে

✅

CPU + RAM proactive forecasting

দুটো resource একসাথে predict করা

✅

Multi-horizon prediction

৫/১০/১৫/৩০ মিনিট — সব horizon-এ naive-কে হারানো

✅

Monte Carlo uncertainty

Confidence interval সহ prediction

✅

Online learning + drift detection

প্রতি ৩ ঘণ্টায় leakage-free retraining

✅

Cost-aware right-sizing

Actual usage থেকে package recommendation

✅

Live production deployment

Real hosting business-এ deployed system