CNN Hyper-Parameter Optimizer based on Evolutionary Selection and GOW Approach for Crimes Tweet Detection

Social media is widely utilized for crime, highlighting the need for automated crime detection and prevention systems that can reach a vast audience. There is substantial study on specific social media crimes, but a holistic model to detect them is needed. This article classifies cybercrime tweets using deep neural networks, specifically deep CNNs, a common computer vision method. Although promising, deep neural networks need further research to find their best topology. Manual iteration and intensive testing limit the deep CNN model's practicality. CNN embedding dimensions, dense unit count, learning rate, and batch size affect model performance. Thus, constructing an effective CNN model with little manual intervention and domain expertise is difficult. This study reduces dimensionality and improves model performance by selecting the most relevant and informative dataset features using CNN hyperparameters. Five CNN feature selection methods using optimal hyperparameters are shown. These approaches use CNN, SO-roulette wheel, SO-tournament, SO-linear rank, SO-exponential rank, and Grey Wolf Optimizer to identify and categorize Arabic criminal tweets. The CNN-SO-tournament model recognizes illicit Arabic tweets with 99.60% accuracy. This study substantially improves Twitter crime prevention.