Is Single Enough? A Joint Spatiotemporal Feature Learning Framework for Multivariate Time Series Prediction

Area of Science:

• Data Mining
• Artificial Intelligence
• Time Series Analysis

Background:

• Multivariate time series prediction (TSP) is crucial but challenging.
• Existing Fuzzy Cognitive Map (FCM) approaches struggle with nonlinear spatiotemporal dependencies, limiting accuracy.
• Single-mode feature extraction is insufficient for complex TSP.

Purpose of the Study:

• To propose a novel joint spatiotemporal feature learning framework for multivariate TSP.
• To enhance the accuracy of multivariate time series forecasting.
• To address the limitations of existing FCM-based methods.

Main Methods:

• Developed a Spatio-Temporal Fuzzy Cognitive Map (STFCM) framework.
• Incorporated a mix-resolution spatial module with sparse autoencoders (SAEs) for feature extraction.
• Utilized a mix-order spatiotemporal module with high-order FCMs (HFCMs) for dynamic modeling.
• Employed batch gradient descent for efficient weight updates.

Main Results:

• The proposed STFCM framework demonstrates superior performance on four real-world datasets.
• Experimental results validate the effectiveness of the mix-resolution spatial and mix-order spatiotemporal modules.
• STFCM achieves higher prediction accuracy compared to state-of-the-art methods.

Conclusions:

• The STFCM framework offers significant advantages for multivariate time series prediction.
• The joint spatiotemporal feature learning approach effectively models complex dependencies.
• STFCM represents a promising advancement in time series forecasting.