Stacked ensemble modeling with personalized weightsbased on clustering

Abstract

Ensemble modeling improves classification model performance and has been widely used at predictive modeling. Among ensemble methods, the stacking technique performs a weighted average over the predictions of individual base models, where each model has a fixed weight. This paper proposes a novel approach to the weight assignment where each data point has different set of weights for base models at stacking. This is achieved by employing first clustering technique K-Means and then training the combiner models for each cluster. To measure the improvements at generalization accuracy, the proposed algorithm has been tested at 22 benchmark datasets while the number of clusters varied from 2 to 12 for each. The algorithm delivered improved performance up to 41% of datasets.