An efficient hybrid Hopfield convolutional neural network for detecting spam bots in Twitter platform

Abstract

Recently, social media platforms have become very popular as they offer unbelievable opportunities to their users. Twitter is one of the social media platforms on which a huge number of people exchange their messages by posting tweets. However, this platform is usually used by automated accounts called bots. Such bots are used to spread fake news, fake ideas, and products. Hence, it is essential to detect the presence of spam bots on Twitter. In order to detect spam bots on Twitter, an effective feature selection technique using a novel hybrid deep learning model is introduced in this paper. This paper proposes a novel spam bot detection system for the Twitter social network that combines profile and tweet-based features. Initially, the Twitter data are pre-processed to improve the accuracy of classification. The pre-processing stage involves various steps such as stopping word removal, tokenization, stemming, n-gram identification, user mention, and vocabulary density and richness. After pre-processing, the tweets are given to the next stage for feature extraction. In this stage, the user profile-based features such as name, screen name, location, and time, as well as the tweet-based features such as hashtags, retweeting of tweets, etc., are extracted from the tweets. The extracted features are then subjected to feature selection, where a meta-heuristic-based optimization algorithm called the Binary Golden Search Optimization algorithm (BGSO) is used. This method helps to reduce the feature dimensionality and overfitting issues. In order to improve the optimization algorithm’s searching ability, an X-shaped transfer function is used. Finally, the selected features are provided to the novel Hybrid Hopfield Dilated Depthwise Separable Convolutional Neural Network (HHD²SCNN) based classification model, where the output layer classifies the given tweets as spam bots or legitimate. The proposed method is experimentally verified, and the performance metrics are evaluated. Simulation is done using the Python tool, and the Cresci 2017 dataset is used. Simulation results show that the proposed HHD²SCNN model provides better accuracy, having an accuracy of 98.40 % compared to the existing techniques. Also, the proposed hybrid deep learning model achieved a precision of 98.40 %, recall of 98.40 %, specificity of 98.40 %, F-score of 98.40 %, and kappa of 96.80 %. Thus, the proposed technique achieves better results compared to the existing techniques.