A0734
Title: Multi-task learning for compositional data based on sparse network lasso regularization
Authors: Akira Okazaki - The University of Electro-Communications (Japan) [presenting]
Shuichi Kawano - The University of Electro-Communications (Japan)
Abstract: A network lasso enables us to construct a model for each sample, which is known as multi-task learning. It is used in various fields of research, in particular life science in which, the obtained data contain heterogeneity that varies among samples. In such a case, general models that are common to all samples fail to extract the effective information, which is related to heterogeneity. On the other hand, compositional data, which consist of the proportions of a composition, are used in the fields of geology and life science for microbiome analysis. Existing methods for multi-task learning cannot be applied to compositional data due to their intrinsic properties. In this research, we propose a multi-task learning method for compositional data using a sparse network lasso regularization. We focus on a symmetric form of the log-contrast model, which is a regression model with compositional covariates. The symmetric form is extended to the locally symmetric form in which each sample has a different regression coefficient vector. These regression coefficient vectors are clustered by the network lasso regularization and selected by the L1-norm regularization. The effectiveness of the proposed method is shown through simulation studies and application to gut microbiome data.
