A Novel Method to Calculate Sample Weights for FCM Regression

Regression is an important prediction method to establish models between variables. The primitive regression algorithms ignore the sample weights, and consider all samples play an equal role in regression. But this kind of algorithms often loses efficacy when dealing with outliers, since outliers disturb the regression models greatly. For traditional switching regression, sample membership varies with models when sample weights are equal. In this paper, we propose an adaptive sample weighting method for FCM regression, in which sample membership and sample weights are computed simultaneously. Such method can make outlier sample weights as small as possible. Numerical experiments suggest that our approach is effective.