Clustering with varying risks of false assignments in discrete latent variable model

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Latent variable models are common for analyzing unlabeled data structure in clustering.
• Model-based clustering typically minimizes overall false assignment errors.
• Existing methods lack flexibility in handling differential error importance across clusters.

Purpose of the Study:

• Introduce a novel clustering rule for differential error control.
• Develop a method to estimate the false assignment rate in clustering.
• Address limitations of existing methods in handling varied error sensitivities.

Main Methods:

• Introduce the concept of false assignment rate for clustering.
• Utilize an extended likelihood approach for estimating the false assignment rate.
• Propose VRclust, a new clustering rule designed for differentiated error management.

Main Results:

• Demonstrate the estimation of false assignment rate using real data examples.
• Show that VRclust effectively controls various errors differently across clusters.
• Simulation studies confirm the consistency of error controls with increasing sample size.

Conclusions:

• VRclust provides a flexible framework for managing differential errors in clustering.
• The proposed method enhances the applicability of model-based clustering.
• The false assignment rate estimation and VRclust offer improved control in data analysis.