Abstract:

 According to the euclidean distance, the original LLE (locally linear embedding) algorithm chooses the neighborhood. If the data was originated from multiple classes,  the correct neighborhood relationship could not be obtained. In order to solve this problem, an improved MLLE(modified LLE) was proposed. In MLLE algorithm, the distance matrix was modified, which could make the distance longger between classes and smaller within classes, and so could make the neighborhood in one class as far as possible. The test of Chinese text clustering showed that the MLLE algorithm could improve the clustering visualization and the recognition rate.

Key words: manifold learning, LLE algorithm, MLLE algorithm, Chinese text clustering