Abstract
Making an accurate prediction of an unknown system only from a short-term time series is difficult due to the lack of sufficient information, especially in a multi-step-ahead manner. However, a high-dimensional short-term time series contains rich dynamical information, and also becomes increasingly available in many fields. In this work, by exploiting spatiotemporal information (STI) transformation scheme that transforms such high-dimensional/spatial information to temporal information, we developed a new method called MT-GPRMachine to achieve accurate prediction from a short-term time series. Specifically, we first construct a specific multi-task GPR which is multiple linked STI mappings to transform high dimensional/spatial information into temporal/dynamical information of any given target variable, and then makes multi step-ahead prediction of the target variable by solving those STI mappings. The multi-step-ahead prediction results on various synthetic and real-world datasets clearly validated that MT-GPRMachine outperformed other existing approaches.
Introduction
Predicting the future state of complex dynamical systems is a fundamental challenge across scientific disciplines, from ecology and biology to economics and engineering. Traditional time series forecasting methods, such as Autoregressive Integrated Moving Average (ARIMA) or Long Short-Term Memory (LSTM) networks, typically rely on extensive historical datasets to learn the underlying dynamics. However, in many practical scenarios, researchers only have access to "snapshot" data—very short time series that lack the temporal depth required for conventional training.
This paper addresses the critical problem of multi-step-ahead prediction from limited temporal data. While temporal depth may be lacking, modern observational techniques often provide high-dimensional "spatial" data, measuring numerous variables simultaneously. The proposed MT-GPRMachine framework leverages this high-dimensional spatial information to compensate for short-term temporal observations, effectively reconstructing system dynamics through a novel spatiotemporal information (STI) transformation.
Methodology
The MT-GPRMachine framework operates on two primary pillars: Spatiotemporal Information (STI) Transformation and Multi-task Gaussian Process Regression (MT-GPR).
Spatiotemporal Information (STI) Transformation
Based on delay embedding theory, the method establishes a mapping $\Psi(X(t)) = Y(t)$ that relates the high-dimensional observed spatial data $X(t)$ to the temporal information of a target variable $Y(t)$. By organizing these mappings into a Hankel matrix structure—where each row represents a time-shifted version of the dynamics—the framework creates a mathematically linked set of tasks. This structure allows the model to "virtually expand" the available data by treating future time steps as related learning objectives rather than independent variables.
Multi-task Gaussian Process Regression (MT-GPR)
Instead of solving each prediction step in isolation, MT-GPRMachine treats them as linked tasks. The multi-task GPR shares information across these tasks, constrained by the physical and dynamical relationships inherent in the STI matrix. Specifically, the model solves tasks simultaneously subject to the constraint $\Psi_l(X(m)) = \Psi_{l+1}(X(m-1))$, ensuring temporal consistency. This "cooperation" between linked GPRs helps cancel out individual prediction errors and provides significant robustness against noise.
Experimental Evaluation
The researchers evaluated MT-GPRMachine against 13 baseline models, including LSTM, ARIMA, Support Vector Regression (SVR), and the Anticipated Learning Machine (ALM). Testing was conducted on a diverse range of datasets:
- Synthetic Systems: A 90-dimensional time-variant coupled Lorenz system and a 64-dimensional coupled pendulum system.
- Environmental Data: Longitude and latitude prediction for Typhoon Marcus, and oxygen concentration forecasting in the Northern Gulf of Mexico (Plankton dataset).
- Climate and Infrastructure: Ground-level temperature prediction in Texas (Ozone dataset) and amplitude/frequency prediction for a HUAWEI wireless channel dataset.
Results and Discussion
Experimental results demonstrated that MT-GPRMachine consistently outperformed existing approaches in both accuracy and robustness. On the Plankton dataset, the model achieved a reduction in Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) of over 65% compared to the next-best baseline.
As shown in the comparative analysis, MT-GPRMachine exhibits exceptional noise immunity. In synthetic tests with Gaussian noise ($N(0, 3)$), it significantly outperformed the ALM, which is often considered a standard for embedding-based methods but typically struggles with stochastic interference. Furthermore, the model proved robust to missing data; in wireless channel tests, it maintained a Pearson Correlation Coefficient (PCC) above 0.8 even when 10% of the training data was randomly omitted.
Despite these strengths, the authors noted that GPR-based methods are sensitive to high-frequency fluctuations. For highly non-smooth data, such as the Ozone dataset, a moving average filter was required to achieve optimal performance. Additionally, the computational complexity of Multi-task GPR increases with the number of tasks and dimensions, suggesting a need for more efficient implementations in extremely high-dimensional settings.
Conclusion
MT-GPRMachine represents a significant advancement in time series forecasting from "small data." By substituting temporal depth with spatial breadth through STI transformation and enforcing dynamical constraints via multi-task learning, the framework enables accurate multi-step-ahead predictions where traditional models fail. Future work will focus on developing hybrid models to better handle non-smooth data and optimizing computational efficiency for large-scale applications.
Summary updated