Nearest Neighbor Multivariate Time Series Forecasting.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Multivariate time series (MTS) forecasting is crucial in many fields.
• Spatial-temporal graph neural networks (STGNNs) are popular but limited by computational complexity and inability to use full datasets.
• Existing methods struggle with sparse, discontinuous correlations over long periods, yielding minor improvements.

Purpose of the Study:

• To develop a novel MTS forecasting framework that overcomes limitations of current STGNNs.
• To enable models to access and leverage entire datasets for improved pattern recognition.
• To enhance forecasting accuracy by identifying sparse, similar patterns across variables and time.

Main Methods:

• Introduced a nearest neighbor MTS (NN-MTS) forecasting framework.
• Utilized a nearest neighbor retrieval mechanism over a large datastore of cached series.
• Developed a hybrid spatial-temporal encoder (HSTEncoder) for capturing long-term temporal and short-term spatial-temporal dependencies.

Main Results:

• NN-MTS demonstrated significant improvements in forecasting performance on real-world datasets.
• The framework effectively extracts sparse, distributed, yet similar patterns across multiple variables.
• NN-MTS provides direct access to the entire dataset at test time without additional training.

Conclusions:

• NN-MTS offers a highly expressive and efficient approach to MTS forecasting.
• The framework shows superior interpretability and efficiency, enhancing its application prospects.
• NN-MTS opens new avenues for leveraging large datasets in MTS modeling.