Research On Feature-based Inductive Transfer Leanring Algorithm

Transfer learning is a new area in machine learning, Feature-based transfer learningmethod is one of many transfer technologies, it aims to find a more complete feature spacebetween the source domain and the target domain, and transfer the feature between thesource domain and the target domain. In the inductive transfer learning problems, we oftenencounter a situation that there are little labeled data and a large number of unlabeled databoth in source domain and target domain, how to get the same or similar feature spacewhich get from some labeled data among a large number of unlabeled data become the keyof solving these kind of problems.This thesis presents a feature-based meta-prior model. The model can extract a largenumber of feature from target domain unlabeled data, and fix the sub-feature weight valuedepending on the labeled data in source domain. Then dimensions reducing by making useof these prior messages into the labeled data in target domain and finally generate a morecomplete meta-model with good transfer performance. Experiments show the FIMP’saccuracy and efficiency has been improved in the classification.