A semi-supervised classification method based on transduction of labeled data

The semi-supervised classification problem with partially labeled data is very important in the research area of pattern recognition and machine learning. In this paper, an approach based on transduction of labeled data is proposed to improve current classification methods. The general knowledge about the attribute of data distribution is used to carry out transduction. Employing this kind of knowledge, the commonly existent mode of the distribution corresponding to each labeled sample can be effectively found by mean shift, and the data at the mode can be regarded as having the same label with the original labeled sample with high confidence. Using the mode data instead of the original labeled data for classification can be capable of improving classification performance. Encouraging experimental results both on synthetic data and real-world handwritten characters validate the applicability and effectiveness of the approach.