N-CovSel, a new strategy for feature selection in N-way data

Area of Science:

• Data Science
• Chemometrics
• Multivariate Data Analysis

Background:

• Variable selection is crucial in data analysis, but many methods are limited to 2-way data matrices.
• Extending feature selection to higher-order (N-way) structures is challenging due to difficulties in assessing variable relevance in multi-way contexts.

Purpose of the Study:

• To propose and evaluate a novel feature selection strategy, N-CovSel, specifically designed for N-way data structures.
• To demonstrate the capability of N-CovSel in identifying relevant features in both simulated and real-world datasets.

Main Methods:

• The study introduces N-CovSel, a feature selection strategy based on the Covariance Selection (CovSel) approach, adapted for N-way data.
• The N-CovSel method allows for the selection of features with varying dimensionality (1-way up to (N-1)-way).
• The method's performance was validated using a simulated dataset to verify feature selection against known ground truth and a real-world dataset.

Main Results:

• N-CovSel successfully identified meaningful features in both simulated and real datasets, demonstrating its effectiveness for N-way data.
• The method proved capable of selecting features compatible with the underlying structure of the data.

Conclusions:

• N-CovSel offers a robust solution for feature selection in complex N-way datasets, addressing a gap in current methodologies.
• The study also proposed further analysis strategies for selected features, including sequential multi-block methods and N-way Partial Least Squares (N-PLS) based approaches.