A New Domain Adaptation Method Based on Rules Discovered from Cross-Domain Features

摘要

Traditional classification methods in machine learning assume that training data and testing data should share the same feature space and have the same data distribution. In real world applications, however, this assumption often does not hold. If there are very few labeled instances in the target domain for training, it is time-consuming to label them manually. In this case, a source domain which has semantic relationships with the target domain but has the different feature space or distribution can be used to assist the classification. In this paper, we propose a new method using rules to help the domain adaptation, which can well represent the knowledge relationships between source domain and target domain. In this algorithm we first discover term-term rules according to the term relationships in target domain to build the knowledge bridge, then we reconstruct the source domain using these rules and get a better classifier to improve the cross-domain classification performance. We conduct several cross-domain data sets and demonstrate that the proposed method is easy to understand and it has a better performance compared to state-of-art transfer algorithms.