Talks and presentations

Mode merging for non-Gaussian Finite Mixtures

September 02, 2022

Talk, MBC2, Catania, Italy

Finite mixture models can be interpreted as a model representing heterogeneous sub-populations within the whole population. However, more care is needed when associating a mixture component with a cluster, because a mixture model may fit more components than the number of clusters. Modal merging via the mean shift algorithm can help identifysuch multi-component clusters. So far, most of the related works are focused on the Gaussian finite mixture. As the non-Gaussian finite mixtures gain attention, the need to address the component-cluster correspondence issue in these mixture models grows. In light of this issue, we introduce mode merging methods for several non-Gaussian finite mixtures including power-exponential, normal variance mixture and normal variance-mean mixture.

Why would I need this again?

July 07, 2021

Talk, Statistics and Actuarial Science Student Seminar Series, Waterloo, Canada

(Spoiler: I needed it again.) We encounter in our courses numerous esoteric concepts and techniques, and wonder why we learn them. Once the final exam or project leaves out hand, we bid farewell to most of the course content and move on. We don’t need them again, after all. (Un)fortunately, my research journey so far has proven otherwise. In this talk, I will share this experience in the context of my most recent problem, which is mixture component merging.

Topics in component merging in model-based clustering

June 10, 2021

Talk, 2021 Annual Meeting of the Statistical Society of Canada (SSC 2021), Canada

Model-based clustering models heterogeneous populations within a data set, where each component is often viewed as a cluster. However, the number of components may not always match that of the underlying clusters. In the case of over-estimation, merging some mixture components could help with identifying more informative clusters. In light of this issue, we will discuss some topics related to component merging in model-based clustering.

Adapting the Motivated Strategies for Learning Questionnaire (MSLQ) to help students take control of their learning

June 02, 2021

Talk, Society for Teaching and Learning in Higher Eduacation (STLHE 2021), Canada

Finding a regularized common projection in model-based clustering (…or so I thought)

March 17, 2021

Talk, Statistics and Actuarial Science Student Seminar Series, Waterloo, Canada

Suppose that a finite mixture model identified 3 components from a data set. Which variables contributed the most in identifying said components? That was the initial question. In particular, I wanted to find a smaller, rather than larger, number of important variables, since that would make the investigator’s life easier. Contrary to my expectation, the first-devised strategy was not satisfactory, so I had to backtrack a little after much confusion. On that note, I will discuss briefly the problem of finding a regularized common projection for a finite mixture model, and share my experience in backtracking and re-evaluating my approach.

Effect of penalisation on a mixture of factor analysers

December 15, 2019

Talk, 12th International Conference ofthe ERCIM WG on Computational and Methodological Statistics (CMStatistics 2019), Waterloo, Canada

Factor analysers can be used to obtain a parsimonious estimate of component-wise covariance matrices in a finite mixture model. In addition, one could achieve further parsimony in estimated covariance matrix by penalising on the factor loading matrix. However, an increasing magnitude of penalisation coefficient may result in degenerate factor loading estimates, which may have an adverse effect on maximum likelihood estimation of model parameters. To this end, we investigate the effect of penalisation on sparse estimation of parameters in a finite mixture of factor analysers. We also investigate the effect of such estimates in model-based clustering settings.

Subspace clustering for the finite mixture of generalized hyperbolic distributions

September 05, 2018

Talk, International Workshop on Model-based Clustering and Classification (MBC2), Catania, Italy

The finite mixture of generalised hyperbolic distributions is a flexible model for clustering, but its large number of parameters for estimation, especially in high dimensions, can make it computationally expensive to work with. In light of this issue, we provide an extension of the subspace clustering technique developed for finite Gaussian mixtures to that of generalised hyperbolic distribution. The methodology will be demonstrated with numerical experiments.