B0796
Title: Ensemble clustering for learning mixtures of Gaussian processes
Authors: Mimi Zhang - Trinity College Dublin (Ireland) [presenting]
Xiantao Zhao - Trinity College Dublin (Ireland)
Emmanuel Akeweje - Trinity College Dublin (Ireland)
Abstract: An ensemble clustering framework is developed to efficiently identify functional data's latent cluster labels from a Gaussian process mixture. The approach exploits the independence and Gaussianity of the coefficients in the Karhunen-Loeve expansion of a Gaussian random function. In the framework, each base clustering in the ensemble is obtained by fitting a univariate Gaussian mixture model to the projection coefficients of the functional data onto one basis function, with the different basis functions being orthonormal. The computational complexity for identifying the cluster labels is much lower than that of state-of-the-art methods, and theoretical guarantees are provided on the identifiability and learnability of Gaussian process mixtures. Extensive experimentation on synthetic and real datasets validates the superiority of the method over existing techniques. To facilitate application, the framework is implemented into a comprehensive Python package available on GitHub.
