A study on classification of small sample based on stochastic configuration networks

Due to the limited availability of samples and the high cost of annotation, small sample classification presents significant challenges. Traditional models often struggle with poor generalisation and inadequate inter-class separability. To tackle these issues, this paper introduces an ensemble model called RS-SCN, which combines stochastic configuration networks (SCN) with the random subspace (RS) method. In this approach, the feature space is partitioned into random subspaces, each used to train an independent SCN model. The outputs of these models are then integrated through majority voting. This strategy reduces dependence on large-scale datasets while enhancing generalisation performance. Experimental results demonstrate that RS-SCN surpasses traditional methods in both generalisation and robustness. As a result, this approach enhances the applicability of SCN to both small-sample classification and function approximation tasks, offering an effective solution to the generalisation challenges posed by limited data.