Caformer: Rethinking Time-Series Forecasting From Causal Perspective.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Causal Inference

Background:

• Time-series forecasting is crucial but challenged by nonstationary data and environmental confounders.
• Existing methods struggle to disentangle true temporal patterns from spurious correlations caused by external factors.

Purpose of the Study:

• To introduce Caformer, a novel framework for time-series forecasting based on causal reasoning.
• To effectively capture cross-dimension and cross-time dependencies while mitigating environmental influences.

Main Methods:

• Caformer utilizes four modules: dynamic learner (cross-dimension dependencies), temporal learner (causal cross-time dependencies).
• Environment learner and decompose learner extract environmental factors and apply backdoor adjustment to correct for confounding effects.

Main Results:

• Caformer achieves state-of-the-art performance in both long-term and short-term forecasting.
• Demonstrated significant Mean Squared Error (MSE) reductions: 26.2% on traffic and 21.8% on electricity datasets compared to PatchTST.
• Achieved top rankings on the M4 dataset for short-term forecasting.

Conclusions:

• Caformer effectively addresses confounding environmental factors in time-series forecasting.
• The framework offers superior predictive accuracy and provides interpretable insights into learned dependencies.