迁移学习算法研究.ppt

Finally, I want to introduce how to construct the target classifiers. The are two ways. First, combine the output source classifiers based on the new representation. Second, Combine all the source domain data and train a unified classifier using supervised learning algorithms, e.g., SVM and Logistic Regression. The two accuracies are denoted as CRAv and CRAu, respectively. * The image data used in the paper by Luo, which appears in CIKM’08 is adopted to evaluate our algorithm. There are two top categories flower and traffic, and the top categories are used for classification. Every top category contains four sub-categories. To construct one domain data, we randomly selected one sub-categories from the two top categories, then we have four domains. After that, we randomly select one domain as target domain, we can totally have 96 three-source vs one-target domain problems. * The data set sentiment classification is also used in this paper, and we construct 4 three-source vs one-target domain problems. The accuracy on target domain data is used as the evaluation measure. Both SVM and LR are used to train classifiers on new the representation. * The baselines include: The supervised learning algorithms on original features: SVM and Logistic Regression. The Embedding method based on autoencoders (EAER) The Marginalized Stacked Denoising Autoencoders (mSDA) The Transfer Component Analysis (TCA) The Transfer learning from multiple sources (CCR3) The algorithms mSDA TCA and CCR3 are transfer learning methods. Our methods are CRAv and CRAu. For the methods which can not handle multiple sources, we train the classifiers from each source domain and merged data of all sources (r+1 accuracies). Finally, maximal, mean and minimal values are reported. * This slide shows the Results on 96 image classification problems. From these results, we find that our algorithm achieve the best results. 1) CRA is significantly better than the traditional machine learning algorit