(AAAI26) XLinear: A Lightweight and Accurate MLP-Based Model for Long-Term Time Series Forecasting with Exogenous Inputs
This is the official implementation of XLinear: A Lightweight and Accurate MLP-Based Model for Long-Term Time Series Forecasting with Exogenous Inputs.
We have designed an extremely simple and efficient time series model—XLinear—based on MLP and sigmoid to handle real-world forecasting tasks with exogenous inputs, bridging the gap between efficiency and accuracy in time series forecasting.
XLinear consists solely of two sets of gating modules with identical structures, which are designed to filter out noisy features in the temporal and variable dimensions, enhance critical features, and strengthen temporal patterns within the time series, respectively. To mitigate information interference between different dimensions, we draw on the approach proposed in TimeXer for learning global representations of endogenous variables, thereby facilitating information integration across these two dimensions.
First, we conduct forecasting tasks with exogenous variables on 7 commonly used datasets. For this scenario, we designate the last variable as the endogenous variable, with the remaining variables serving as exogenous variables.
Second, we supplement 5 additional datasets with strong exogenous factor interference for comparative experiments. To comprehensively evaluate the model's performance in hydrological forecasting scenarios, we incorporate new metrics such as NSE, KGE, and MAPE to assess its effectiveness.
Furthermore, we investigate the performance of XLinear in multivariate forecasting scenarios.
In addition to achieving exceptional accuracy, XLinear maintains remarkably high efficiency, reaching a level comparable to that of DLinear and RLinear. We investigated the efficiency of XLinear in both multivariate forecasting scenarios and univariate forecasting scenarios with exogenous inputs. Although DLinear and RLinear outperform XLinear by a marginal advantage in terms of efficiency, their predictive accuracy is considerably inferior. In contrast, compared with Transformer-based models that achieve higher accuracy, XLinear exhibits an approximate 30% improvement in training speed while consuming less GPU memory.
Furthermore, we investigate the capability of XLinear to learn temporal patterns from longer lookback windows.
Concurrently, we compare it with several lightweight and high-precision time series models in terms of variations in model resource consumption and running speed as the lookback window expands.
1.Datasets can be obtained via the following links: ETT Weather Electricity Traffic Crop DO_409202 DO_425012
2.Install Pytorch and other necessary dependencies.
pip install -r requirements.txt3.All dataset scripts are centralized in the script folder. Execute the following startup commands in the main directory. Examples are as follows:
bash ./script/multi_forcasting/etth1.shIf you have any questions or concerns, please contact us at {Warren.Jin@csiro.au, Yhuang@mail.hzau.edu.cn, Zaiwen.Feng@mail.hzau.edu.cn} or submit an issue.