LGA (ICLR 2026)

This repository is the official PyTorch implementation of LGA: Local Geometry Attention for Time Series Forecasting under Realistic Corruptions

Local Geometry Attention

Standard attention uses one global metric for all queries. LGA learns a query-specific distance metric via local Gaussian process theory — each query adapts to its own local data geometry.

  

(Left) Two query points produce distinct attention surfaces that match their local cluster shapes. (Right) Standard attention is distracted by an anomaly (high score on red dot); LGA ignores it.

TSRBench

The first robustness benchmark for time series forecasting. Uses Extreme Value Theory (EVT) to inject realistic corruptions at 5 severity levels.

Want to inject noise into your own datasets? TSRBench is available as a standalone pip-installable toolkit with 6 corruption types, a Python API, and support for any CSV/DataFrame input. Check out the TSRBench repository for details.

Original → Exponential Spike → Level Shift → Combined

Getting Started

Requirements

• Python >= 3.9
• CUDA-compatible GPU (recommended)

Install the required packages:

pip install -r requirements.txt

Data Preparation

1. Download the datasets from Autoformer.
2. Create a folder named ./dataset in the root directory of this project.
3. Place all downloaded files and folders within the ./dataset folder.

The expected directory structure:

dataset/
├── ETT-small/
│   ├── ETTh1.csv
│   ├── ETTh2.csv
│   ├── ETTm1.csv
│   └── ETTm2.csv
├── electricity/
│   └── electricity.csv
└── weather/
    └── weather.csv

Generating Corrupted Benchmarks (TSRBench)

Before training or evaluation, generate the corrupted test datasets first:

bash ./TSRBench/generate_noise.sh

This reads from the original dataset files and generates corrupted versions at 5 severity levels:

dataset/ETT-small/
├── ETTh1.csv                          # original (unchanged)
├── ETTh1_noise/
│   ├── ETTh1_level_1_type_shift.csv
│   ├── ETTh1_level_1_type_spike.csv
│   ├── ETTh1_level_1_type_combined.csv
│   ├── ...
│   └── ETTh1_level_5_type_combined.csv
├── ETTh2_noise/
│   └── ...

See TSRBench/README.md for detailed documentation on noise injection and custom datasets.

Training

We provide scripts for all datasets used in the paper. For example:

# ETTh1
bash ./scripts/ETTh1.sh

# Weather
bash ./scripts/weather.sh

# Electricity
bash ./scripts/electricity.sh

You can find more scripts in the ./scripts folder for other datasets.

Robustness Evaluation

To evaluate a trained model under noise corruptions:

python -u run.py \
  --is_training 0 \
  --root_path ./dataset/ETT-small/ \
  --data_path ETTh1.csv \
  --model_id ETTh1_336_96 \
  --model PatchTST \
  --data ETTh1 \
  --features M \
  --seq_len 336 \
  --pred_len 96 \
  --enc_in 7 \
  --lga \
  --itr 1

Project Structure

.
├── run.py                 # Main entry point for training and evaluation
├── models/
│   └── PatchTST.py        # Model wrapper with LGA integration
├── layers/
│   ├── PatchTST_backbone.py   # Core architecture including LGA module
│   ├── PatchTST_layers.py     # Supporting layers and embeddings
│   └── RevIN.py               # Reversible Instance Normalization
├── exp/
│   ├── exp_basic.py       # Base experiment class
│   └── exp_main.py        # Training, validation, and testing logic
├── data/
│   ├── data_factory.py    # Data provider
│   └── data_loader.py     # Dataset classes
├── utils/
│   ├── metrics.py         # Evaluation metrics (MSE, MAE, etc.)
│   ├── tools.py           # EarlyStopping, learning rate scheduling, etc.
│   ├── timefeatures.py    # Temporal feature engineering
│   └── masking.py         # Masking utilities
├── scripts/               # Shell scripts for reproducing experiments
└── TSRBench/              # Time Series Robustness Benchmark toolkit

Citation

If you find this repo useful for your research, please cite our paper:

@inproceedings{
kim2026local,
title={Local Geometry Attention for Time Series Forecasting under Realistic Corruptions},
author={Dongbin Kim and Youngjoo Park and Woojin Jeong and Jaewook Lee},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=NCQPCxN7ds}
}

Acknowledgements

We would like to express our appreciation for the following repositories, which provided valuable code bases and datasets:

• PatchTST
• Autoformer
• Time-Series-Library

Contact

If you have any questions or want to use the code, please contact dongbin413@snu.ac.kr