<p>The Stochastic Configuration Network (SCN) is a single-hidden-layer feedforward neural network with a unique supervision mechanism for configuring hidden node parameters. This mechanism endows the model with the potential for unlimited learning and rapid convergence. Due to its high learning efficiency and adaptive adjustment of random parameter ranges, SCN has seen rapid development in recent years. However, existing SCN models and their variants primarily focus on supervised learning tasks. In practical scenarios, labeled data are often scarce and difficult to obtain, while unlabeled data are abundant. Training SCNs with only a small amount of labeled data and a large amount of unlabeled data tends to lead to overfitting on the limited labeled samples. To enhance the generalization ability of SCNs in semi-supervised learning, this paper proposes a Semi-Supervised Classification Model based on SCN (SS-SCN). First, SS-SCN integrates L₂ and manifold regularization to mitigate overfitting. The Laplacian matrix introduced in the manifold regularization enables SS-SCN to model the intrinsic geometric structure of both labeled and unlabeled data, thereby enhancing its capability for semi-supervised learning. Furthermore, a class weight matrix is introduced to address class imbalance in classification tasks. Finally, a novel supervision mechanism is designed within the SS-SCN framework, and its universal approximation property under this constraint is theoretically proved. Overall, experimental results on multiple datasets present a comparison between SS-SCN and existing algorithms (such as SCN, DeepSCN, L₂SCN, SS-RVFL, and LPSCN) across binary and multiclass semi-supervised classification tasks.</p>