B1584
Title: Multi-task learning regression based on convex clustering
Authors: Akira Okazaki - Kyushu University (Japan) [presenting]
Shuichi Kawano - Kyushu University (Japan)
Abstract: When related multiple datasets are observed, it is expected to the existence of some common information among them. Multi-task learning (MTL) is a methodology that aims to improve the estimation and prediction of multiple models set for each dataset by sharing common information. In the MTL, each model is called a task. One of the natural assumptions in the practical situation is that tasks are classified into some clusters with their characteristics. In the framework of MTL for regression models, the group fused regularization approach performs clustering tasks by shrinking the difference of regression coefficients among tasks. The approach enables the transfer of common information within the same cluster and improves the estimation of the regression coefficients. However, the approach also transfers the information between different clusters, which worsens the estimation and prediction. An MTL method is proposed with a centroid parameter representing a cluster centre of the task. Because the proposed method separates parameters into the parameters for regression coefficients and the parameters for clustering, it can improve estimation and prediction accuracy for regression coefficients. The effectiveness of the proposed method is shown through Monte Carlo simulations and applications to real data.