Robust Bilinear Probabilistic PCA Using a Matrix Variate t Distribution.

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Bilinear Probabilistic Principal Component Analysis (BPPCA) is a dimension reduction technique for matrix data.
• Standard BPPCA relies on Gaussian assumptions, making it sensitive to outlying observations.
• Robustness is crucial for real-world datasets that often contain outliers.

Purpose of the Study:

• To develop a robust extension of BPPCA that can handle outlying matrix-valued observations.
• To introduce a new model, t-BPPCA, based on the matrix variate t-distribution.
• To provide efficient algorithms for parameter estimation and outlier detection.

Main Methods:

• Developed t-BPPCA by incorporating a matrix variate t-distribution, which includes a robustness tuning parameter.
• Introduced a hierarchical representation using a Gamma distributed latent weight variable.
• Designed two accelerated expectation-maximization (EM)-like algorithms for parameter estimation.

Main Results:

• Experiments on synthetic and real datasets demonstrate that t-BPPCA is more robust and accurate than existing methods, especially in the presence of outliers.
• The t-BPPCA model effectively identifies outliers using expected latent weights.
• Outlier detection with t-BPPCA proved more reliable than its vector-based counterpart.

Conclusions:

• The proposed t-BPPCA offers a robust and accurate alternative to standard BPPCA for dimension reduction on matrix data.
• The matrix variate t-distribution provides enhanced resilience against outliers.
• t-BPPCA facilitates reliable outlier detection, a significant advantage in data analysis.