Simultaneous Application of PLS Regression and FCM-type Clustering

Although fuzzy c-regression models (FCRM) is a useful tool for switching regression, it often suffers from collinearity problems because of the inherent sparsity of samples. This paper proposes a new approach to switching PLS regression that performs local PLS regression in conjunction with data partitioning by fuzzy clustering. In the proposed method, the PLS regression part plays a role for estimating lower dimensional latent variables that are useful for prediction of some external criteria while the clustering part is responsible for unsupervised classification of samples considering local data structure. The characteristic features are revealed in several numerical experiments including feature extraction from POS transaction data.