The-AnkitPatel/netpredict-telemetry-lgbm

⚡ NetPredict: High-Speed GBDT Latency Forecaster with Conformal Coverage

Model Description

NetPredict is an optimized LightGBM Gradient Boosted Decision Tree (GBDT) regressor designed for microsecond-scale point latency forecasting in high-concurrency cloud networks and distributed telemetry pipelines.

When paired with Split Conformal Prediction, the model provides distribution-free, finite-sample coverage guarantees ($1 - \alpha = 0.95$) over non-linear queuing delays and bufferbloat anomalies.

• Author: Ankit Kumar Patel (The-AnkitPatel)
• Affiliation: National Institute of Technology, Uttarakhand | IIT Roorkee Telemetry Testbed
• Open Source Role: Upstream Contributor @ Cloud Native Computing Foundation (CNCF OpenTelemetry)
• Framework: LightGBM GBDT + Scipy Conformal Calibration
• Inference Latency: $0.14\text{ms}$ ($p99$ SLA)

---

Intended Use & Capabilities

• eBPF-driven flow scheduling: Predict buffer saturation 5–15 minutes before tail-drop packet loss.
• Uncertainty-aware microservices: Yields exact $[\hat{y} - q, \hat{y} + q]$ intervals with guaranteed $95.0%$ empirical test set coverage.

---

Performance & Evaluation Metrics

Metric	Holdout Test Value	Unit
RMSE	0.3982	milliseconds
MAE	0.2814	milliseconds
$R^2$ Score	0.9924	dimensionless
95% Conformal Coverage	95.12%	empirical guarantee satisfied
$p99$ Inference Latency	0.142	milliseconds

---

Quickstart & Inference

import pandas as pd
import lightgbm as lgb
from huggingface_hub import hf_hub_download

# Download model weights from Hugging Face Hub
model_path = hf_hub_download(repo_id="The-AnkitPatel/netpredict-telemetry-lgbm", filename="model.txt")
bst = lgb.Booster(model_file=model_path)

# Example telemetry ingress vector
sample_flow = pd.DataFrame([{
    'ingress_mbps': 650.0,
    'active_flows': 220,
    'arrival_kpps': 580.0,
    'buffer_util_pct': 74.5,
    'loss_rate_pct': 0.4,
    'jitter_ms': 3.2
}])

latency_point_pred = bst.predict(sample_flow)
print(f"Predicted One-Way Queuing Latency: {latency_point_pred[0]:.3f} ms")

---

Live Interactive Space

Test the model live with interactive sliders at:
👉 Hugging Face Space: NetPredict Conformal Telemetry