Fake Comment Detection Based on Generative Adversarial Networks

Fake comments are a widespread problem in various online activities. This issue not only impacts the consumer experience and service quality but also makes it difficult to distinguish genuine information from the vast pool of comments. Therefore, automated detection of fake comments has become crucial, and this challenging task has garnered considerable attention. Machine learning and deep learning have offered several effective approaches for automatically detecting fake comments, with deep learning, in particular, significantly advancing the field of fake comment detection. However, deep learning often requires a substantial amount of labeled data as support, and the lack of high-quality labeled data has long been a major obstacle in fake comment detection. To address the issue of insufficient labeled data, generative adversarial networks (GANs) have been introduced to compensate for the shortage by generating high-quality fake samples. This method utilizes a limited number of labeled samples and generated fake samples for adversarial learning, thereby enhancing the performance of fake comment detection methods. In this approach, a specific discriminator network structure is designed, combined with long short-term memory networks (LSTM), to optimize feature propagation between network layers. This approach not only maximizes the learning of contextual features but also helps alleviate the potential performance degradation that deep networks may encounter. Ultimately, the proposed method is evaluated using two publicly available high-quality fake comment datasets. The results show that this method achieves a high validation accuracy of 97.87% and an average validation accuracy of 91.04%. Furthermore, metrics such as F1-score, recall, and precision also outperform other comparative methods, with average values of 91.26%, 92.4%, and 95.65%, respectively. A comprehensive analysis of various evaluation metrics suggests that this method outperforms other approaches.