Transfer Learning Algorithms Based On Maximum Entropy Model

Most of traditional machine learning and data mining algorithms are based on the assumption that the training and test data must be in the same feature space and follow the same distribution. Therefore, they more focus on a single learning task which is usually independent of other learning tasks or previous knowledge. However, in real applications, the feature space and data distributions change frequently. In such case, we have to re-collect and re-label large amounts of data for training a new model in the traditional learning framework, which is very expensive and time-consuming.Transfer learning, as a new learning framework, aims at building a system with the ability to recognize and apply knowledge and skills learned in previous tasks to novel tasks. Under the settings of inductive transfer learning and transductive transfer learning, two algorithms based on the maximum entropy model are proposed. One is the weighted algorithm of inductive transfer learning, named WTLME, which focuses on transferring one single-source domain to target domain. The other is the ensemble algorithm of transductive transfer learning, named SFEC, which focuses on transferring multi-source domain to target domain.WTLME obtains the target classifier model with high accuracy by transferring the model parameters from source domain to target domain and adjusting the instance weights of target domain. Extensive experiments show the efficiency of WTLME on the Web page and review datasets. To further improve the accuracy of sentiment classification on the review datasets, SFEC are proposed, which trains the ensemble classifier on the mixed data from labeled source domain data and unlabeled target domain data, and then pre-labels the target domain data and feed back the ones with higher reliability into the ensemble classifier for the next iterative training. Experiments on the review datasets also show that not only the accuracy of sentiment classification improved greatly by SFEC, but also the negative transfer phenomenon is alleviated.