Tensor analysis of animal behavior by matricization and feature selection.

Area of Science:

• Neuroscience
• Data Science
• Bioinformatics

Background:

• Neurobehavioral research generates complex multi-dimensional tensor (MDT) data.
• Analyzing MDT data is crucial for understanding neural circuitries but poses interpretation challenges.
• Standard multivariate analysis tools are designed for 2-dimensional tensor (2DT) data.

Purpose of the Study:

• To investigate the impact of different matricization and feature selection methods on multivariate analysis of MDT neurobehavioral data.
• To optimize the transformation of MDT data into 2DT for enhanced analysis.
• To identify robust methods for distinguishing behavioral phenotypes using transformed data.

Main Methods:

• Applied various matricization techniques (Index Construction, Feature Concatenation) to MDT data.
• Employed filter and embedded feature selection methods to identify informative 2DT features.
• Utilized cross-validation and holdout validation with multiple classifiers to assess performance on zebrafish visual-motor response data.

Main Results:

• Feature Concatenation combined with embedded feature selection or union operations yielded optimal performance for classifiers.
• The optimized approach revealed unique behavioral differences between wildtype and visually-impaired mutant zebrafish.
• These subtle differences were not discernible through standard MDT or multivariate analysis alone.

Conclusions:

• Matricization and feature selection are effective strategies for analyzing complex MDT neurobehavioral data.
• The Feature Concatenation method with embedded feature selection offers a powerful approach for enhancing multivariate analysis.
• This methodology provides novel insights into behavioral phenotypes and their underlying neural mechanisms.