Probabilistic Rank-One Tensor Analysis With Concurrent Regularizations.

Area of Science:

• Multilinear tensor analysis
• Machine learning
• Dimensionality reduction

Background:

• Subspace learning for tensors is gaining traction, with multilinear extensions of Principal Component Analysis (PCA) and Probabilistic PCA (PPCA).
• Existing methods like Tucker and CANDECOMP/PARAFAC (CP) based PPCA have limitations: Tucker models have restrictive subspace representation and rotational ambiguity, while CP models are prone to overfitting.

Purpose of the Study:

• To introduce Probabilistic Rank-One Tensor Analysis (PROTA), a novel CP-based multilinear PPCA.
• To address the limitations of existing tensor subspace learning methods, specifically rotational ambiguity and overfitting.
• To enhance subspace representation flexibility and improve robustness in tensor data analysis.

Main Methods:

• PROTA utilizes a CP-based multilinear PPCA framework with a more flexible subspace representation than Tucker-based methods.
• Introduced Concurrent Regularizations (CRs) to mitigate overfitting in CP-based PPCAs by adjusting noise variance or latent feature moments.
• Developed a Bayesian treatment of PROTA for automatic feature determination and enhanced robustness against overfitting.

Main Results:

• PROTA demonstrates superior performance in subspace estimation and classification tasks compared to existing methods.
• Concurrent Regularizations (CRs) effectively alleviate overfitting issues in CP-based tensor analysis.
• The Bayesian approach to PROTA provides automatic feature selection and improved generalization.

Conclusions:

• PROTA offers a more flexible and robust approach to tensor subspace learning.
• Concurrent Regularizations are effective strategies for improving the performance of CP-based tensor analysis.
• The proposed Bayesian PROTA framework enhances automatic feature determination and robustness, making it a powerful tool for analyzing complex tensor data.