Tensor analysis of animal behavior by matricization and feature selection.

Area of Science:

• Neuroscience
• Behavioral Science
• Data Science

Background:

• Neurobehavioral research generates complex multi-dimensional tensor (MDT) data.
• Standard multivariate analysis (MVA) struggles with high-dimensional MDT data.
• Existing dimensionality reduction methods like matricization may lose information or introduce noise.

Purpose of the Study:

• To systematically evaluate matricization and feature selection methods for MDT data analysis.
• To assess the impact of these methods on downstream MVA performance.
• To identify optimal strategies for analyzing neurobehavioral MDT datasets.

Main Methods:

• Applied various Index Construction and Feature Concatenation matricization techniques to MDT data.
• Utilized filter and embedded methods for informative 2D tensor (2DT) feature selection.
• Evaluated performance using a zebrafish visual-motor response dataset (wild-types vs. mutants) via classification tasks and cross-validation.

Main Results:

• The combination of Feature Concatenation matricization and embedded feature selection yielded the best MVA performance across most classifiers.
• This approach revealed unique behavioral distinctions between wild-type and visually-impaired zebrafish, missed by standard analyses.
• Specific matricization and feature selection strategies significantly impact the effectiveness of downstream MVA.

Conclusions:

• Matricization coupled with feature selection is a powerful approach for analyzing complex neurobehavioral MDT data.
• The optimal strategy involves Feature Concatenation and embedded feature selection for enhanced MVA.
• This methodology enhances the ability to uncover subtle behavioral phenotypes and underlying neural circuitry.