abhishek8aiml/METR-LA

---
license: mit
task_categories:
- time-series-forecasting
- tabular-regression
tags:
- traffic-prediction
- time-series
- graph-neural-networks
- transportation
size_categories:
- 1M<n<10M
---

# METR-LA Traffic Dataset

## Dataset Description

This dataset contains traffic flow data for time series forecasting tasks, commonly used with Graph Neural Networks and specifically the Diffusion Convolutional Recurrent Neural Network (DCRNN) model.

## Dataset Structure

### Data Format

- **Format**: Parquet files for efficient loading and analysis
- **Splits**: train (70%), validation (10%), test (20%) - **temporal splits** preserving chronological order
- **Features**: Time series traffic flow data with temporal and spatial dimensions

### Split Strategy

- **Temporal splitting**: Data is split chronologically to prevent data leakage
- **All sensors included**: Each split contains data for all sensors at each time step
- **Training period**: Earliest 70% of time samples across all sensors
- **Validation period**: Next 10% of time samples across all sensors
- **Test period**: Latest 20% of time samples across all sensors
- **Graph structure preserved**: Spatial relationships maintained in all splits

### Data Schema

- `node_id`: Sensor/node identifier (0-206 for METR-LA, 0-324 for PEMS-BAY)
- `t0_timestamp`: ISO 8601 timestamp of the reference time point (t+0) for each sequence
- `x_t*_d*`: Input features at different time offsets and dimensions
- `x_t-11_d0` to `x_t+0_d0`: Traffic flow values at 12 historical time steps
- `x_t-11_d1` to `x_t+0_d1`: Time-of-day features (normalized 0-1)
- `y_t*_d*`: Target values at future time steps and dimensions
- `y_t+1_d0` to `y_t+12_d0`: Traffic flow predictions for next 12 time steps
- `y_t+1_d1` to `y_t+12_d1`: Time-of-day features for prediction horizon

### Dataset Statistics

- **Total time series samples**: ~34K (METR-LA) / ~52K (PEMS-BAY)
- **Total records**: ~7M (METR-LA) / ~17M (PEMS-BAY)
- **Records per sample**: 207 (METR-LA) / 325 (PEMS-BAY) sensors
- **Temporal resolution**: 5-minute intervals
- **Prediction horizon**: 1 hour (12 time steps)

## Usage

```python
from datasets import Dataset, DatasetDict
import pandas as pd

# Load from local parquet files
train_df = pd.read_parquet("METR-LA/train.parquet")
val_df = pd.read_parquet("METR-LA/val.parquet")
test_df = pd.read_parquet("METR-LA/test.parquet")

ds = DatasetDict({
"train": Dataset.from_pandas(train_df, preserve_index=False),
"val": Dataset.from_pandas(val_df, preserve_index=False),
"test": Dataset.from_pandas(test_df, preserve_index=False)
})

print(f"Train records: {len(ds['train']):,}")
print(f"Val records: {len(ds['val']):,}")
print(f"Test records: {len(ds['test']):,}")
```

## Citation

If you use this dataset, please cite the original DCRNN paper:

```bibtex
@inproceedings{li2018dcrnn_traffic,
title={{Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting}},
author={{Li, Yaguang and Yu, Rose and Shahabi, Cyrus and Liu, Yan}},
booktitle={{International Conference on Learning Representations}},
year={{2018}}
}
```

## Dataset Generation

The code used to generate this Hugging Face-compatible dataset can be found at [witgaw/DCRNN](https://github.com/witgaw/DCRNN), a fork of the original DCRNN repository with enhanced data processing capabilities.

## Original Data Source

This dataset is derived from the original METR-LA dataset used in the DCRNN paper.

## License

MIT License - See the [original repository LICENSE](https://github.com/liyaguang/DCRNN/blob/master/LICENSE) for details.