The Fisher discriminant analysis(FDA) is a classical supervised dimensionality reduction method in statistical pattern recognition. The FDA can maximize the scatter between different classes, while minimizing the scatter within each class, but the analysis process of the FDA only utilizes the labeled data and ignores the unlabeled data. Therefore, a semi-supervised local fisher method based on probability class and uncorrelated discriminant(SLFisher), which enables the data pairs in different classes to be separated from each other and the nearby data pairs in the same class to be closed after dimensionality reduction. Two benchmark datasets are applied in the experiment, and the results show that the SLFisher can greatly improve recognition rate.