This paper presents a method for semi-supervised multi-class classification based on matrix factorization. The method called semi-supervised online kernel matrix factorization (SS-OKMF) performs a semantic embedding by finding a non-linear mapping to a low-dimensional semantic space. An important characteristic of the SS-OKMF method is that the new low-dimensional semantic representation can be learned in a semi-supervised fashion. Therefore, the annotated instances can be used to maximize the discrimination between classes, but also, the non-annotated instances can be exploited to estimate the intrinsic manifold structure of the data. The non-linear modeling is based on kernel methods with a learning-in-a-budget strategy that allows keeping low memory requirements. This strategy, along with an online formulation based on stochastic gradient descent, reduces the computation time and keeps low computational requirements in large-scale problems. According to the experimental evaluation performed on several datasets of different nature (i.e. images, biosignals, and synthetic data), SS-OKMF, in comparison with several non-linear supervised, unsupervised and semi-supervised dimensionality reduction methods, presents a competitive performance in classification tasks under a transductive learning setup preserving a lower computational cost.