Semi-Supervised learning is a hot topic of machine learning recently, relative to supervised learning and unsupervised learning methods, a semi-supervised learning algorithm can make use of labeled samples and unlabeled samples in the same time, obviously, in this form the learning algorithm can get a better performance. Semi-supervised learning can be divided into two ways :Semi-supervised classification and semi-supervised clustering. A semi-supervised clustering algorithm use some supervised samples to instruct the clustering procedure in order to get a better result, and a semi-supervised classification use some unlabeled samples to train a better classify. This paper based on the semi-supervised learning and pays its attention on semi-supervised clustering algorithms and semi-supervised dimension reduction.The primary work of this paper can be summarized as follows:(1) Base on the ROC, a semi-supervised algorithm (called Semi-ROC) is proposed, with the help of supervised samples, the new algorithms can get a better performance, the experiment result on the artificial dataset and UCI dataset shows the proposed algorithm is effective and feasible.(2)Several semi-supervised kernel clustering algorithms including semi-supervised kernel fuzzy -c means (SKFCM) and semi-supervised kernel possibilistic -c means (SKPCM) was proposed. The result on the artificial and uci dataset shows its efficiency. We also test the SKFCM algorithm in an image segmentation experiment, with the help of supervised information, a better segment result is got.(3) Proposed a semi-supervised dimension reduction algorithm, based on the algorithm, an image retrieval method use the semi-supervised dimension reduction algorithm is proposed and implemented. The test results on Corel dataset show this method is effective.Based on the research work (3), a model system for automatic image retrieval is implemented under the IDE Visual C++6.0. The system consists of real-time image retrieval and image information access. |