A Multi-domain Adaptation for sentiment classification algorithm based on Class Distribution

At present, most multi-domain adaptation for sentiment classification algorithms use all source domains to train the classifier with no selecting and dynamic dealing with the different source domains. This will result in those source domains very dissimilar to target have negative impact on domain adaptation. In this paper, we propose Multi-domain Adaptation algorithm based on the Class Distribution (MACD). First, the information of class distribution is used to select some adaptive base classifiers from all source domains. Then add the ‘self-labeled’ samples into training data, in which, the selection of samples is dynamically adjusted with the similarity between the source and target domain. Last, the final ensemble classifier is constructed using the information of class distribution. The experimental results have shown that the MACD algorithm is effective and superior to some existing approaches in accuracy.