A Hybrid Deep Learning Architecture for Social Media Bots Detection Based on BiGRU-LSTM and GloVe Word Embedding

Social media platforms have opened avenues for communication, information sharing, and engaging with others online. Automated accounts, known as social media bots, have been observed to engage in harmful activities such as disseminating misinformation, participating in online propaganda and election interference, spreading spam, cyberbullying, and harassing people. This paper proposes a new hybrid architecture based on semantic word embedding and Recurrent Neural Networks (RNNs) to detect social media bots. The research methodology includes the use of Global Vectors (GloVe) for text representation to convert tweets into vectors and combining the Bidirectional Gated Recurrent Units (BiGRU) and Long Short-Term Memory (LSTM) algorithms for semantic text-based classification. Using the proposed architecture, the training process was conducted with two datasets, Cresci-2017 and Twibot-20. The effectiveness of the approach in detecting automated accounts was assessed using five evaluation metrics: Precision, Accuracy, Recall, and F1-score. The proposed approach showed outstanding results in identifying social media bots based only on text-based content, achieving a Precision of 100%, Accuracy of 99.73%, Recall of 99.56%, and F1-Score of 99.63% using the Twibot-20 dataset. Moreover, the proposed architecture surpassed the results obtained by the state-of-the-art approach and showed resilience to overfitting and the ability to detect social media bots effectively in unseen and recent data. This highlights the importance of utilizing deep learning methods and semantic word representations to effectively address issues related to detecting and managing social media bot operations.