CMStatistics 2015: Start Registration
View Submission - CMStatistics
B0767
Title: Clusterwise multiblock PLS regression Authors:  Stephanie Bougeard - ANSES (France)
Gilbert Saporta - CNAM (France)
Ndeye Niang - CNAM (France) [presenting]
Abstract: Clusterwise regression methods aim at partitioning data sets into clusters characterized by their specific coefficients in the regression model. Usually, one dependent variable is linearly related to independent variables which are in a single data table. We present clusterwise multiblock PLS: an extension of clusterwise PLS regression to multiresponse variables and independent variables organized in meaningful blocks. This block structure is taken into account through a set of weights based on the importance of the block on the response prediction. This new method provides a partition of the data such as each of its cluster is associated with its own PLS model, which is then used to improve the overall fit of the prediction step. To do so, a new observation is first assigned to the relevant cluster minimizing a specific distance measure or maximizing the class membership probability. The prediction is then performed using the associated local model or using model averaging strategies. This general approach is based on a clear criterion to minimize and can be directly extended to other multiblock regression methods. The clusterwise multiblock PLS regression will be illustrated on both synthetic and real data.