Hussain Basha S
Computer Science
December 2023
There has been a rise in academic and professional interest in how to identify and eliminate dangerous social bots operating within social networks. The widely utilised machine learning-based approach to bot detection results in an unbalanced distribution of samples across categories. As a result of classifier bias, minority samples are rarely identified. Because of this, we propose an enhanced conditional generative adversarial network (enhanced CGAN) to enlarge unbalanced data sets prior to applying training classifiers to enhance the precision with which social bots may be detected. We propose a modified clustering approach, the Gaussian kernel density peak clustering algorithm (GKDPCA), to construct an auxiliary condition, as it prevents the formation of data augmentation noise and removes inequalities in distributions between and within social bot classes. We also introduce the Wasserstein distance with a gradient penalty to the CGAN convergence judgement condition to fix the model collapse and gradient disappearance issues of the original CGAN. In this experimental study, we evaluate three widely used oversampling techniques. Oversampling is analysed in terms of the degree of imbalance and the expansion ratio of the original data, with the enhanced CGAN outperforming the others. Enhanced CGAN outperforms three commonly used oversampling algorithms in terms of F1-score, G-mean, and AUC in experimental data.
955- 972
© Airo Journals 2021
