Product Centred Dirichlet Processes for Bayesian Multiview Clustering.

Area of Science:

• Statistics
• Machine Learning
• Computational Biology

Background:

• Bayesian clustering methods are well-established, but multi-view clustering remains underdeveloped.
• Modeling statistical dependence between clusterings from different data views presents significant challenges.
• Existing methods struggle with the complexity of partition spaces, limiting the modeling of cross-view dependencies.

Purpose of the Study:

• To develop a novel Bayesian framework for multi-view clustering that explicitly models dependencies between clusterings.
• To introduce a new prior, CLustering with Independence Centring (CLIC), for analyzing distinct yet dependent clusterings across multiple data views.
• To provide a computationally efficient and theoretically sound method for multi-view clustering analysis.

Main Methods:

• The proposed CLustering with Independence Centring (CLIC) prior is introduced, based on the product centered Dirichlet process (PCDP).
• Theoretical properties of the CLIC model are derived, including marginal and joint partition distributions.
• A marginal Gibbs sampler is developed for efficient posterior computation, alongside a finite approximation to prove accuracy.

Main Results:

• CLIC successfully models the dependence between clusterings across different views using a single parameter.
• The method accurately characterizes view-specific partitions while providing inference on the level of dependence.
• Performance was validated on synthetic data and in an epidemiological application.

Conclusions:

• CLIC offers a robust and effective solution for Bayesian multi-view clustering.
• The framework accurately captures both individual view clusterings and their interdependencies.
• This approach advances the analysis of complex, multi-modal datasets in various scientific domains.