Social media analysis for criminal behavior detection using CNN-RNN modelling

Social networks have become an essential source of data to recognize criminal behavior and require advanced approaches for automatic learning for a precise classification. This research presents a methodology based on learning in depth for the recognition of criminal behavior using a hybrid folding network (CNN) and a long-term memory model (LSTM). The study uses an Olid data registration, a reference corpus for the detection of offensive languages ​​to form and evaluate the proposed framework. The preliminary treatment includes the TF-IDF vectorization (Term Frequency-Inverse Document Frequency) with a long-term sub-linear frequency scale, a minimum document of the document of 5, a maximum of 10,000 features, standard L2 and a N-gramma of (1.2) to improve the extraction of the characteristics. The component of the CNN records the local space characteristics of the text representations, while the LSTM components moderate sequential addictions and improve the understanding of the context. The experimental results show an accuracy of the classification of 93%, indicating the solidity of the model when determining the criminal intention in publications on social networks. The results underline the effectiveness of learning deeply in the automation of the crimes and constitute a basis for future polls in the legal studies controlled by artificial intelligence. The proposed executive has the potential to support the security authorities in surveillance and intervention in real time and to improve public security and IT security.