Probabilistic Linear Discriminant Analysis Based on L₁-Norm and Its Bayesian Variational Inference.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• Probabilistic linear discriminant analysis (PLDA) is widely used for feature extraction in supervised learning.
• The standard PLDA model uses the squared L2-norm, implicitly assuming Gaussian noise, which can be sensitive to outliers.
• Real-world data often exhibit non-Gaussian noise, limiting the effectiveness of standard PLDA.

Purpose of the Study:

• To propose a robust PLDA model (L1-PLDA) that accommodates non-Gaussian noise distributions, specifically Laplacian noise.
• To enhance outlier detection capabilities within the PLDA framework.
• To improve classification performance on datasets with potential outliers or non-Gaussian noise.

Main Methods:

• Developed a novel L1-PLDA model by assuming a Laplacian noise distribution for model errors.
• Introduced a latent variable to represent the Laplacian distribution as a mixture of Gaussian distributions.
• Employed the variational expectation-maximization (EM) algorithm for parameter learning.

Main Results:

• The proposed L1-PLDA model demonstrated superior performance compared to standard PLDA.
• Experiments confirmed the effectiveness of L1-PLDA in both classification and outlier detection tasks.
• The introduced latent variable proved useful for identifying data outliers.

Conclusions:

• L1-PLDA offers a robust alternative to standard PLDA, particularly for datasets with non-Gaussian noise and outliers.
• The model enhances feature extraction by providing better outlier handling and improved classification accuracy.
• The developed method shows significant potential for applications requiring robust pattern recognition.