关键词: 迁移学习, 欧式距离, 余弦距离, 保护型, 归一化约束, 过维数

Abstract: Former researchers commonly study transfer learning algorithms and traditional machine learning from the point of the rationality of formulas, while neglecting the integrality of the problem. As a result, their algorithms are usually unable to thoroughly practice classification when they are applied to specific text classification problem. Via observing the whole process of text classification, it uses cosine distance in k-mean method and gets obviously better results. It proposes protection-type iteration idea. It abandons traditional word feature space and chooses hidden space as the feature vector space and implements normalization constraints. Taking CCI algorithm as an example, this idea is used to create an improved algorithm which is nominated PCCI. This algorithm can prominently raise the classification accuracy of transfer learning, meanwhile reducing the computing complexity. It proves the superiority of the improved algorithm by comparing with other former transfer learning cases through program testing on the database of 20-NewsGroups and Reuters-21578.

Key words: transfer learning, Euclidean distance, cosine distance, protection-type, normalization constraints, over dimension