GLinear: A Linear Predictor for Multivariate Time Series Forecasting

This repository is the official Pytorch implementation of GLinear Predictor: "Paper Link".

Description

The architecture of GLinear predictor is composed of two fully-connected layers of same input size having a Gaussian Error Linear Unit (GeLU) nonlinearity in-between them. Different configurations of input and layers sizes are tested to lead to this final architecture.

The Reversible Instance Normalization (RevIN) is applied to the input and output layer of GLinear model. The normalization layer transforms the original data distribution into a mean-centred distribution, where the distribution discrepancy between different instances is reduced. This normalized data is then applied as new input to the used predictor and then final output is denormalized at last step to provide the final prediction.

1. RevIN Normalization:
2. Linear transformation:
3. GELU activation:
4. Second Linear transformation:
5. RevIN Denormalization:

Acknowledgment

This code is built on the code base of LTSF-Linear Predictors (NLinear, DLinear). We would like to thank the following GitHub repository for their valuable code bases, datasets and detailed descriptions:

https://github.com/cure-lab/LTSF-Linear

Required Environment and libraries can be installed using instructions given in above link.

Furthermore, we would also like to thank the pytorch implementation of RLinear predictor used for comparative analysis, available at: https://github.com/plumprc/RTSF/blob/main/models/RLinear.py

Training the GLinear and Other Predictors:

First, create the conda environment and install the required tools and libraries, as follows:

conda create -n GLinear python=3.6.9
conda activate GLinear
pip install -r requirements.txt

Datasets

All processed datasets are in the ./dataset directory Etth1, Electricity, Weather, and Traffic

Use Following Command to Train Linear Predictors for Varying LookBack Windows:

bash scripts/EXP-LookBackWindow_\&_LongForecasting/Linear_LookBackWindow.sh

Use Following Command to Train Linear Predictors for Varying Prediction Lengths:

bash scripts/EXP-LookBackWindow_\&_LongForecasting/Linear_LongForecasting.sh

Results

All the results will be shown in ./logs directory

Citing

If you use or find this code repository useful, consider citing it as follows:

  title={Bridging Simplicity and Sophistication using GLinear: A Novel Architecture for Enhanced Time Series Prediction},
  author={Syed Tahir Hussain Rizvi; Neel Kanwal; Muddasar Naeem; Alfredo Cuzzocrea; and Antonio Coronato}, % Replace with your actual name
  year={2025},
  eprint={arXiv:2501.01087}, % Replace with the actual arXiv identifier
  publisher={arXiv}}