Related Experiment Videos
Identifying critical variables of principal components for unsupervised feature selection
1School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore. ekzmao@ntu.edu.sg
Summary
This study introduces two algorithms to link principal components (PCs) back to original measurements, improving the interpretability of unsupervised dimensionality reduction. These methods efficiently select relevant features for clearer data analysis and acquisition.
Area of Science:
- Data Science
- Machine Learning
- Bioinformatics
Background:
- Principal Component Analysis (PCA) is a common method for unsupervised dimensionality reduction.
- Principal Components (PCs) lack clear physical meaning, hindering interpretation and data acquisition.
- Existing methods for feature selection in PCA can be computationally intensive.
Purpose of the Study:
- To develop algorithms that link principal components (PCs) back to original measurements.
- To enhance the interpretability of PCA results.
- To reduce computational complexity in feature selection for PCA.
Main Methods:
- Developed two novel algorithms to identify subsets of original measurements that best represent sample projections onto principal axes.
- Evaluated feature subsets based on their capacity to reproduce PCA projections.
- Compared computational complexity with existing data structural similarity-based methods.
Main Results:
- The new algorithms successfully link principal components to original measurements.
- Feature selection based on reproducing PCA projections is effective.
- Significantly reduced computational complexity compared to previous approaches.
Conclusions:
- The developed algorithms provide a method to interpret physically meaningless principal components.
- These algorithms facilitate clearer data acquisition by identifying essential original measurements.
- The approach offers a computationally efficient alternative for feature selection in PCA.