Recursive Exponentially Weighted N-way Partial Least Squares Regression with Recursive-Validation of Hyper-Parameters...

Area of Science:

• Machine Learning
• Data Science
• Signal Processing

Background:

• High-dimensional, multi-way (tensor) data present challenges for traditional regression models.
• Adaptive modeling of complex processes requires efficient real-time algorithms.
• Existing recursive algorithms may suffer from information loss or suboptimal hyper-parameter tuning.

Purpose of the Study:

• To introduce a novel tensor-input/tensor-output Recursive Exponentially Weighted N-Way Partial Least Squares (REW-NPLS) regression algorithm.
• To develop an adaptive modeling approach for complex processes using high-dimensional tensor data in real-time.
• To propose a Recursive-Validation method for hyper-parameter estimation, enhancing recursive algorithms.

Main Methods:

• The proposed REW-NPLS algorithm combines Recursive Exponentially Weighted PLS with tensor-based methods.
• The algorithm is designed for tensor-input/tensor-output data, ensuring no information loss.
• A novel Recursive-Validation procedure is implemented for hyper-parameter estimation, replacing conventional cross-validation.

Main Results:

• The REW-NPLS algorithm demonstrated efficient and robust treatment of high-dimensional tensor data.
• Performance was validated against state-of-the-art methods using electrocorticography (ECoG) and magnetoencephalography (MEG) datasets.
• The implemented software is suitable for real-time operation, showing promise for Brain-Computer Interface applications.

Conclusions:

• The REW-NPLS algorithm offers a powerful tool for adaptive modeling of complex processes with multi-modal data structures.
• The Recursive-Validation method provides an effective alternative for hyper-parameter estimation in recursive modeling.
• The proposed approach has broad applicability beyond neuroscience, including fields requiring real-time adaptive modeling of complex data.