A1234
Title: Advancements in finite mixture models and flexible model-based clustering techniques
Authors: Samyajoy Pal - LMU Munich (Germany) [presenting]
Christian Heumann - LMU Munich (Germany)
Abstract: The aim is to explore the advancements of modelling multivariate data with finite mixture models and model-based clustering by enhancing flexibility, parameter estimation, and performance across diverse data types. A Dirichlet mixture model (DMM)-based clustering method that utilizes a modified hard EM algorithm and a soft EM algorithm is introduced. The approach outperforms popular clustering algorithms, as demonstrated on both simulated and real-world datasets. Additionally, an alternative parametrization of the Dirichlet distribution is proposed using mean and precision parameters, improving interpretability and estimation accuracy. Innovative estimation techniques, such as Stirling's and moment approximations, provide closed-form solutions that boost model identifiability and computational efficiency, especially in high-dimensional settings. Traditional mixture models are further extended by allowing combinations of identical and non-identical distributions, including mixtures of multivariate skew normal and multivariate generalized hyperbolic distributions. This generalized framework effectively captures complex data structures and accurately identifies underlying patterns. Overall, a robust and flexible toolkit is offered for finite mixture modeling and clustering, advancing the capacity to handle complex data scenarios with improved accuracy and interpretability.
