Multi-source Domains Transfer Learning Algorithm Research

The traditional machine learningmust satisfy that the train data and test data must be the same distribution. If the distribution of test dataset have changed, we need to re-train the predicted classifier to accommodate these changes. The most different of Transfer Learning method is that it can use the knowledge of related domains to help achieve target tasks to train the learning method. Transfer learning can share or transfer the information between the different domains or tasks. It can make traditional machine learning more flexible to learn. Low cost,high efficiency is the most importance feature of Transfer learning. As a kind of Machine learning, transfer learning is widely used in some fields. Such as computer vision, image classification, sentimental analysis and document classification. The Classification results of Transfer learning rely on the similarity of distribution between the source domains with target domain. As the different distribution from different source domains, multi-source domains transfer learning has been proposed. Multi-source domains transfer learning try to find the suitable samples from every different source domains to do the knowledge transfer so that it can decrease the negative transfer happened. For Multi-source domains transfer learning, in this paper, all works are done as follow:1. To solve the problem we discuss above, in this paper, we present a novel method that using pseudo labeled kernel means matching of multi-source domains transfer learning. During the method of pseudo labeled kernel means matching, it can re-assignment the weight of every source domains’ instance more reasonable. After re-assignment the weight of every source domains’ instance, we use the Laplace Figure to compute the weight of every source domains. And then we have the instance’s weight with transfer feature. So that we can train a better Classifierwith these weight. And we do some experiment to confirm this method work.2. To do some further research, we put the iterativeframe into the method discuss above. In order to have a better performance prediction results, the method try to use last time prediction results to train this time Classifier, and we try to analysis the iteration times and the end of the iteration.3.in this paper, we try to combine the method between SVM and KMM, because it can re-assignment according to the optimal classification plane and minimize the distance between source domains and target domains in feature space.And we do some experiment to confirm this two method work.