Semi-supervised learning has recently emerged as a new paradigm in the machine learning community. It aims at exploiting simultaneously labeled and unlabeled data for classification. We introduce here a new semi-supervised algorithm. Its originality is that it relies on a discriminative approach to semi-supervised learning rather than a generative approach, as it is usually the case. We present in details this algorithm for a logistic classifier and show that it can be interpreted as an instance of the Classification Expectation Maximization algorithm. We also provide empirical results on two data sets for sentence classification tasks and analyze the behavior of our methods.