Exploring Sequential Information in Twitter Spam Detection: A Deep Learning Perspective

With recent technological advancements, Internet users can now easily access social media. However, the number of content polluters, commonly referred to as spammers, has increased significantly over time. Spammers repeatedly broadcast unwanted content on social networking sites to distract users. To mitigate this growing issue, it is essential to develop an efficient and robust method for detecting Twitter spammers. This article focuses on text-based analysis, utilizing several word embedding methods, including Word2Vec, GloVe vectors, normal embedding weight initialization, Glorot initialization, and random embedding weight initialization, to extract vocabulary features. Using these vocabulary features, various deep learning models, such as LSTM, BiLSTM, CNN, and Transformer, are trained to classify spam and non-spam content. According to the experimental findings, the highest accuracy and F1 scores for spam detection were achieved by LSTM with uniform random embedding weight initialization, scoring 78.8% and 74%, respectively. The model assigned embedding weights to all vocabulary words, including unknown words in the tweets, leading to better generalization on new test data.