Tensor envelope mixture model for simultaneous clustering and multiway dimension reduction.

Area of Science:

• Multivariate statistics
• Unsupervised learning
• Data science

Background:

• Tensor data are increasingly common in scientific research and biomedical applications.
• Existing dimension reduction techniques may be inefficient for heterogeneous tensor data.
• Current clustering methods are not designed for tensor-variate samples.

Purpose of the Study:

• To propose a novel Tensor Envelope Mixture Model (TEMM) for simultaneous clustering and multiway dimension reduction of tensor data.
• To address limitations of existing methods in handling complex tensor structures.
• To enhance statistical efficiency and reduce estimation variability in tensor data analysis.

Main Methods:

• Developed a model-based clustering approach incorporating tensor-structure-preserving dimension reduction.
• Utilized an expectation-maximization-type algorithm for parameter estimation.
• Constrained model parameters onto lower-dimensional tensor envelope subspaces.

Main Results:

• TEMM demonstrated encouraging empirical performance in simulation studies.
• The method showed effectiveness in a real-world data application.
• TEMM outperformed existing vector and tensor clustering methods in comparative analyses.

Conclusions:

• The proposed TEMM offers a statistically efficient approach for clustering and dimension reduction of tensor data.
• TEMM effectively handles the heterogeneity and complex structure inherent in tensor datasets.
• This method has significant potential for applications in computational biology, brain imaging, and process monitoring.