Dual-stream interactive networks with pearson-mask awareness for multivariate time series forecasting

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multivariate time series forecasting (MTSF) models often struggle to simultaneously capture intra-series temporal dynamics and inter-series dependencies.
• Existing Transformer-based models are limited by single embedding representations, hindering comprehensive analysis of both time and variate dimensions.
• A significant research gap exists in effectively modeling the complex interplay between multiple variates and their temporal evolution.

Purpose of the Study:

• To propose a novel Dual-Stream Interactive Networks with Pearson-Mask Awareness (DSIN-PMA) for enhanced multivariate time series forecasting.
• To address the limitations of existing methods by developing a model that comprehensively understands inter-series interactions and intra-series changes.
• To improve the accuracy and robustness of MTSF by considering both temporal and variate dimensions effectively.

Main Methods:

• Employed a dual-stream embedding structure, incorporating multivariate embedding and time-step embedding for richer data representation.
• Introduced a two-stream network architecture: a cross-multivariate attention with a pearson-mask module for efficient inter-variate dependency learning and noise reduction, and a time-step attention module for temporal pattern discovery.
• Implemented a cross-dimension consistency learning strategy to enhance feature representation and model robustness.

Main Results:

• DSIN-PMA demonstrated significant performance improvements over baseline models across 11 real-world datasets.
• Achieved substantial gains of 5.12%-17.43% compared to state-of-the-art (SOTA) methods in MTSF tasks.
• In-depth analysis confirmed the superiority of the comprehensive dual-stream approach over single-dimensional strategies.

Conclusions:

• The proposed DSIN-PMA model effectively addresses the limitations of existing MTSF methods by jointly modeling temporal and inter-variate dependencies.
• The dual-stream architecture with pearson-mask awareness offers a more robust and accurate solution for complex multivariate time series forecasting.
• The findings highlight the importance of considering both variate and time-step dimensions for superior performance in MTSF.