Abstract

Meituan Bench (MTB) is an enterprise-level benchmarking tool designed for time-series forecasting in real-world business scenarios. Built upon an open-source dataset derived from 10,000 real-world services across various business units, MTB provides a standardized evaluation framework for time-series prediction models. The dataset includes 200 representative services, capturing diverse traffic patterns essential for assessing forecasting performance. Accurate time-series prediction is crucial for optimizing cloud-based elasticity scaling and resource allocation, especially in dynamic environments such as food delivery, where proactive scaling is necessary during traffic surges. MTB establishes performance evaluation standards that align with enterprise requirements, ensuring a fair comparison of forecasting algorithms using real-world traffic data. By offering a comprehensive benchmarking platform, MTB facilitates the development and deployment of more reliable and efficient time-series prediction models for large-scale business applications.

Instructions:

Quickstart

You can download the raw data and use it, or you can directly install mtb through pip to use it

Install MTB: You can install the MTB package directly from PyPI using:
```
pip install mtbenchmark
```
Load Data: MTB implements the PyTorch dataset interface, and you can load data using the following code. For more details, refer to example.py:

The data will be automatically downloaded from Google Drive on the first run, or you can manually download it: link.
```
train_dataset = mt_dataset.MTDataset(features='svc1', split="train")
val_dataset = mt_dataset.MTDataset(features='svc1', split="val")
test_dataset = mt_dataset.MTDataset(features='svc1', split="test")
```

Dataset Files

Files have not been uploaded for this dataset

Datasets

Standard Dataset

MTB: A Time-Series Prediction Benchmarking Tool Tailored to Enterprise Scenarios

Abstract

Quickstart

Dataset Files

QUESTIONS?