A Multiattention-Based Supervised Feature Selection Method for Multivariate Time Series.

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Feature selection is crucial for multivariate time series (MTS) prediction, requiring identification of relevant variables and delays.
• Nonlinear and noisy systems pose significant challenges for traditional MTS feature selection.

Purpose of the Study:

• To propose a multi-attention-based supervised feature selection method for MTS prediction.
• To address the challenges of variable and delay selection in complex, noisy systems.

Main Methods:

• A novel multi-attention mechanism is employed, translating feature weight generation into a bidirectional attention problem.
• Two parallel attention modules process 1D data slices from orthogonal directions, generating dimension-specific weights.
• A global weight generation method combines these weights via dot product for a holistic feature perspective.
• Final feature weights are stabilized using training set statistics to mitigate noise and feature duplication.

Main Results:

• The proposed method demonstrated superior performance across synthesized, small, medium, and practical industrial datasets.
• It outperformed several state-of-the-art baseline feature selection techniques.

Conclusions:

• The multi-attention-based supervised feature selection method effectively handles variable and delay selection in MTS prediction.
• This approach offers a robust solution for complex, noisy time series data, achieving state-of-the-art results.