Robust Multivariate Time Series Forecasting Against Intraseries and Interseries Transitional Shift.

Area of Science:

• Machine Learning
• Time Series Analysis
• Data Science

Background:

• Real-world multivariate time series (MTS) data exhibit nonstationarity, causing distribution shift that challenges forecasting models.
• Existing methods like adaptive normalization and time-variant modeling have limitations in capturing intraseries/interseries correlations and the root causes of distribution shift.

Purpose of the Study:

• To develop a unified probabilistic graphical model (PGM) to jointly address intraseries/interseries correlations and time-variant distributions in nonstationary MTS forecasting.
• To introduce a neural framework, JointPGM, designed to mitigate the limitations of current approaches for MTS forecasting.

Main Methods:

• JointPGM utilizes Fourier basis functions to learn dynamic time factors.
• It incorporates distinct intraseries and interseries learners to capture temporal and spatial dynamics, respectively.
• Gumbel-softmax sampling and multihop propagation are employed for explicit spatial dynamics modeling.

Main Results:

• JointPGM achieves state-of-the-art (SOTA) forecasting performance on six highly nonstationary MTS datasets.
• The model demonstrates effectiveness and efficiency in handling complex temporal and spatial dynamics.
• Experimental validation confirms the model's ability to capture underlying causes of distribution shift.

Conclusions:

• JointPGM offers a unified framework for nonstationary MTS forecasting by jointly modeling correlations and time-variant distributions.
• The proposed neural framework enhances model expressiveness and interpretability in tackling distribution shift.
• The results highlight the potential of JointPGM for advancing MTS forecasting accuracy and robustness.