Weibo Authorship Identification based on Wasserstein generative adversarial networks

During the past years, authorship identification has played a significant role in the public security area. Recently, deep learning based approaches have been used in authorship identification. However, all approaches based on deep learning require a large amount of original data, while the author only posts a limited number of texts to be used as positive samples. As such, the issue of data missing and class imbalance arises, resulting in the classifier overfitting on small data. Furthermore, this approach can be biased by prior of the data. To address these issues, in this paper, we use Wasserstein generative adversarial networks (WGANs) to generate samples for the positive class which lacks of data and present a novel data augmentation framework for authorship identification. To form an augmented training dataset, we mix the generated new samples with the original features in the dataset. As such, the classifier trained on it will not suffer from overfitting and class imbalance, and thus has substantially improved performance. Using crawled Sina Weibo dataset, we empirically evaluate this data augmentation. The experimental results show that our method has a significant improvement of accuracy by 14% compared with powerful baselines. We further validate its effectiveness via a set of comparison experiments.