The-AnkitPatel/netpredict-telemetry-lgbm

NetPredict: LightGBM Network Telemetry Latency Forecaster

Overview

NetPredict is a compact LightGBM regression model trained to forecast one-way queuing delay from high-frequency network telemetry signals. The model is structured for low-latency inference so it can run alongside packet inspection pipelines and software switches to identify impending bufferbloat before packets are dropped.

When paired with inductive split conformal prediction, the model outputs calibrated prediction intervals with finite-sample coverage guarantees at a requested confidence level (default: 95%).

• Author: Ankit Kumar Patel
• Affiliation: National Institute of Technology, Uttarakhand
• Testbed: Distributed Systems & Telemetry Lab, IIT Roorkee
• License: MIT
• Format: LightGBM Booster (model.txt)

Input Features

The model expects 6 input telemetry features:

1. ingress_mbps (float): Ingress traffic rate in Mbps.
2. active_flows (int): Number of concurrent active TCP flows.
3. arrival_kpps (float): Packet arrival frequency in thousands of packets per second.
4. buffer_util_pct (float): Switch queue buffer occupancy percentage (0-100%).
5. loss_rate_pct (float): Packet drop rate percentage (0-100%).
6. jitter_ms (float): Flow inter-arrival jitter in milliseconds.

Target: latency_ms (float), representing end-to-end queuing and propagation delay.

Performance

Evaluated on a held-out test split of 3,000 samples from the eBPF network telemetry dataset:

Metric	Value	Description
Root Mean Squared Error (RMSE)	0.398 ms	Sub-millisecond error across normal and burst regimes
Mean Absolute Error (MAE)	0.281 ms	Nominal queue tracking
R-squared (R2)	0.9924	Captures non-linear queue saturation curve
95% Conformal Coverage	95.12%	Empirical test set coverage at 1 - alpha = 0.95
Median Inference Latency (p50)	0.14 ms	Single-core CPU evaluation

Quickstart

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

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

# Sample input vector
telemetry_data = 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
}])

prediction = bst.predict(telemetry_data)
print(f"Predicted queuing delay: {prediction[0]:.3f} ms")

Conformal Calibration

To compute uncertainty intervals alongside point predictions:

import numpy as np

# Quantile non-conformity threshold from calibration set (alpha = 0.05)
q_hat = 0.742  # ms

point_pred = bst.predict(telemetry_data)[0]
lower_bound = max(0.0, point_pred - q_hat)
upper_bound = point_pred + q_hat

print(f"Point forecast: {point_pred:.2f} ms")
print(f"95% Conformal interval: [{lower_bound:.2f}, {upper_bound:.2f}] ms")

Related Links

• Dataset: The-AnkitPatel/ebpf-network-telemetry
• Live Interactive Space: The-AnkitPatel/netpredict-conformal-telemetry
• Kaggle Notebook: theankitkumarpatel/conformalized-lightgbm-network-telemetry

Citation

@misc{patel2026netpredict,
  author = {Patel, Ankit Kumar},
  title = {NetPredict: LightGBM Network Telemetry Latency Forecaster},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/The-AnkitPatel/netpredict-telemetry-lgbm}}
}