Regularizers on sets of generalized estimates

This paper proposes a novel method for constructing ensembles of recognition algorithms based on stacking technology, incorporating regularizers to enhance the generalization capability of models. The primary focus is on preventing overfitting through the use of majorizing functions that adjust the margins (offsets) of objects from the class boundary. The study investigates the conditions necessary for the correct separation of objects during the training of both base algorithms and the meta-algorithm. A hierarchical agglomerative feature grouping algorithm is proposed, which forms latent features based on intra-class similarity and inter-class differences. It is demonstrated that margin regularization and the transformation of quantitative features into nominal ones improve recognition accuracy. The results show that choosing appropriate parameters for the majorizing functions minimizes the accuracy gap between the base and meta-algorithms. Key advantages of the proposed method: - flexible feature selection for the meta-algorithm based on a greedy strategy; - unification of measurement scales through feature transformation; - robustness to overfitting due to margin regularization. The author declares no conflicts of interests.

Keywords

Authors

References