Ada-MSHyper: Adaptive Multi-Scale Hypergraph Transformer for Time Series Forecasting

✨ This repo is the official implementation of Adaptive Multi-Scale Hypergraph Transformer for Time Series Forecasting.

1 The framework of Ada-MSHyper

Ada-MSHyper is proposed to promote more comprehensive pattern interactions at different scales, which consist of four main parts: Multi-scale Feature Extraction (MFE) Module, Adaptive Hypergraph Learning (AHL) Module, Multi-Scale Interaction Module, and Multi-Scale Fusion Module. The overall framework of Ada-MSHyper is shown as follows:

2 Prerequisites

• Python 3.8.5
• PyTorch 1.13.1
• math, sklearn, numpy, torch_geometric

3 Datasets && Description

📦 You can download the all datasets from datasets. All the datasets are well pre-processed and can be used directly.

4 Running

4.1 Install all dependencies listed in prerequisites

4.2 Download the dataset

4.3 Training

🚀 We provide the experiment scripts of Ada-MSHyper on all dataset under the folder ./scripts. You can obtain the full results by running the following command:

# Train on ETTh1
sh ./scripts/Long-range/ETTh1.sh
# Train on ETTh2
sh ./scripts/Long-range/ETTh2.sh
# Train on ETTm1
sh ./scripts/Long-range/ETTm1.sh
# Train on ETTm2
sh ./scripts/Long-range/ETTm2.sh
# Train on Traffic
sh ./scripts/Long-range/traffic.sh
# Train on Electricity
sh ./scripts/Long-range/electricity.sh
# Train on Weather
sh ./scripts/Long-range/traffic.sh

or obtain specific results by runinng the following command:

# Train on Electricity
python run_longExp.py -data elect -input_size 96 -predict_step 96 -root_path ./data/Electricity/ -data_path electricity.csv -CSCM Conv_Construct
# Train on ETTh1
python -u run_longExp.py --is_training 1 --root_path ./dataset/ --data_path ETTh1.csv --model_id ETTh1_96_192 --model ASHyper --CSCM Bottleneck_Construct --data ETTh1 --features M --seq_len 96 --pred_len 192 --enc_in 7 --des 'Exp' --itr 1 --batch_size 16 --learning_rate 0.0001
# Train on ETTh2
python -u run_longExp.py --is_training 1 --root_path ./dataset/ --data_path ETTh2.csv --model_id ETTh2_96_96 --model ASHyper --CSCM Bottleneck_Construct --data ETTh2 --features M --seq_len 96 --pred_len 96 --enc_in 7 --des 'Exp' --itr 1 --batch_size 32 --learning_rate 0.0001
# Train on ETTm1
python -u run_longExp.py --is_training 1 --root_path ./dataset/ --data_path ETTm1.csv --model_id ETTm1_96_96 --model ASHyper --CSCM Bottleneck_Construct --data ETTm1 --features M --seq_len 96 --pred_len 96 --enc_in 7 --des 'Exp' --itr 1 --batch_size 8 --learning_rate 0.0001
# Train on ETTm2
python -u run_longExp.py --is_training 1 --root_path ./dataset/ --data_path ETTm2.csv --model_id ETTm2_96_96 --model ASHyper --CSCM Bottleneck_Construct --data ETTm2 --features M --seq_len 96 --pred_len 96 --enc_in 7 --des 'Exp' --itr 1 --batch_size 32 --learning_rate 0.001
# Train on Traffic
python -u run_longExp.py --is_training 1 --root_path ./dataset/ --data_path traffic.csv --model_id traffic_96_96 --model ASHyper --CSCM Bottleneck_Construct --data custom --features M --seq_len 96 --pred_len 96 --enc_in 862 --des 'Exp' --itr 1 --batch_size 16 --learning_rate 0.0001
# Train on Electricity
python run_longExp.py --is_training 1 --root_path ./dataset/ --data_path electricity.csv --model_id elect_96_96 --model ASHyper --CSCM Bottleneck_Construct  --data custom --features M --seq_len 96 --pred_len 96 --enc_in 321 --des 'Exp' --itr 1 --batch_size 16 --learning_rate 0.0001
# Train on Weather
python -u run_longExp.py --is_training 1 --root_path ./dataset/ --data_path weather.csv --model_id weather_96_96 --model ASHyper --CSCM Bottleneck_Construct --data custom --features M --seq_len 96 --pred_len 96 --enc_in 21 --des 'Exp' --itr 1 --batch_size 16 --learning_rate 0.0001

5 Main results

We conduct extensive experiments to evaluate the performance and efficiency of Ada-MSHyper, covering long-range, short-range, and ultra-long-range time series forecasting, including 11 real-world benchmarks and 13 baselines.

🏆 Ada-MSHyper achieves consistent state-of-the-art performance on all benchmarks, covering a large variety of series with different frequencies, variate numbers and real-world scenarios.

5.1 Long-range forecasting

5.1.1 Long-range forecasting under multivariate settings.

5.2 Long-range forecasting under univariate settings.

5.2 Short-range forecasting

5.3 Ultra-long-range forecasting

Citation

😀 If you find this repo useful, please cite our paper.

@inproceedings{shangada,
  title={Ada-MSHyper: Adaptive Multi-Scale Hypergraph Transformer for Time Series Forecasting},
  author={Shang, Zongjiang and Chen, Ling and Wu, Binqing and Cui, Dongliang},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}，
  year={2024}
}

Concat

If you have any questions, please feel free to contact zongjiangshang@cs.zju.edu.cn

Other works

📝 Our other works are shown as follows：

Single-step forecasting: Chen L, Chen D, Shang Z, et al. Multi-scale adaptive graph neural network for multivariate time series forecasting. TKDE, 2023, 35(10): 10748-10761. Code Link

AutoML related forecasting: Chen D, Chen L, Shang Z, et al. Scale-aware neural architecture search for multivariate time series forecasting. TKDD, 2024. Code Link

Long-range time series forecasting: Shang Z, Chen L, Wu B, et al. MSHyper: Multi-Scale Hypergraph Transformer for Long-Range Time Series Forecasting. arXiv, 2024: arXiv: 2401.09261. Code Link

The code and documentation are still being finalized, and the final version will be released after the conference.