Title: Predictive clustering using a component-based approach
Authors: Michio Yamamoto - Kyoto University (Japan) [presenting]
Atsushi Kawaguchi - Saga University (Japan)
Heungsun Hwang - McGill University (Canada)
Abstract: A novel clustering method is proposed to identify a cluster structure that is related to outcome variables and to predict cluster memberships of future individuals based on a large number of explanatory variables. Technically, the proposed method carries out outcome-guided dimension reduction (ODR) and the clustering of the dimension-reduced subspace simultaneously. ODR is proven to coincide with a type of partial least squares (PLS) regression, indicating that the proposed method represents a simultaneous approach to PLS and subspace-clustering. PLS is a component-based approach, in which components are defined as linear combinations of explanatory variables. Thus, the combined method can help a researcher to interpret which explanatory variables have effects on the cluster structure based on the weight in linear combinations. In addition, sparse estimation of weights is adopted to obtain a more interpretable and stable cluster structure. Simulated and real data analyses show that the proposed method can provide a cluster structure that is associated with outcome variables and predict cluster memberships of future individuals well.