A linear feature extraction for multiclass classification problems based on class mean and covariance discriminant...

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Multiclass classification requires effective feature extraction methods.
• Existing methods like Linear Discriminant Analysis (LDA) can overemphasize large distances.
• Combining complementary feature extraction schemes presents challenges in optimizing performance.

Purpose of the Study:

• To propose a novel parametric linear feature extraction method for multiclass classification.
• To integrate discriminant information from class means and covariances effectively.
• To develop a computationally efficient framework that avoids sample-based error estimation.

Main Methods:

• The proposed method combines approximate pairwise accuracy (aPAC) and common-mean feature extraction (CMFE).
• aPAC is used to exploit discriminant information about class means, offering an alternative to LDA.
• A fast spanning-tree-based parametric classification accuracy estimator is developed for feature merging and sorting.

Main Results:

• The integrated approach leverages complementary discriminant information from class means and covariances.
• The parametric nature of the method significantly reduces computational cost compared to sample-based approaches.
• Experimental results demonstrate satisfactory performance on both simulated and real-world data.

Conclusions:

• The proposed parametric linear feature extraction method effectively enhances multiclass classification.
• The integration of aPAC and CMFE, coupled with an efficient accuracy estimator, provides a robust solution.
• This framework offers a computationally advantageous alternative for feature extraction in classification tasks.