CATS: Constructing Auxiliary Time Series for Multivariate Time Series Forecasting
Core Concepts
CATS method enhances multivariate time series forecasting by generating Auxiliary Time Series to represent inter-series relationships.
Abstract
CATS introduces a method to improve Multivariate Time Series Forecasting (MTSF) by constructing Auxiliary Time Series (ATS) from Original Time Series (OTS). The key principles of ATS include continuity, sparsity, and variability. CATS achieves state-of-the-art results with a basic 2-layer MLP as the core predictor. It effectively balances intra-series and inter-series relationships in MTSF. The shifting problem example demonstrates how CATS functions as a 2D temporal-contextual attention mechanism. By incorporating ATS into the prediction of OTS, CATS boosts overall performance. The paper identifies three key principles essential for the efficacy of ATS: continuity, sparsity, and variability. Various types of ATS constructors are proposed to handle diverse MTSF challenges.
CATS
Stats
CATS achieves state-of-the-art results with a basic 2-layer MLP as the core predictor.
The shifting problem example demonstrates how CATS functions as a 2D temporal-contextual attention mechanism.
CATS significantly outperforms existing models in average MSE across various datasets.
Quotes
"CATS empowers time series predictors with efficient capturing of inter-series relationships."
"ATS are crucial when inter-series relationships are weak or challenging to learn."
"CATS reduces complexity and parameters compared to previous multivariate models."