Latent dirichlet allocation based blog analysis for criminal intention detection system

Crimes cause several costs to the society, including direct economic costs, victim costs, and other intangible costs. Recently, the research works of implementing computerized system to address the problem of crime has seen a growing interest. In this work, we propose a criminal intention detection system, which objective is to detect the intention of committing a crime by analyzing the content of text documents from article sources found in the internet. The crime intention can be detected from the collection of documents if the topic of the text is properly categorized. We propose an early-warning system to detect the crime activity intention using latent Dirichlet allocation (LDA) and collaborative representation classifier (CRC). Our proposed system involves two stages. In the first stage, we employed LDA as a feature learning method to extract the representation of documents in the article sources, and for the second stage, we used the extracted features from LDA to construct an overcomplete dictionary for CRC to build a classifier to find the related topic for a new testing document. CRC solves an l2-norm optimization problem to find the related topic for a new testing document. Comparing with l1-norm optimization problem in sparse representation classifier (SRC), l2-norm in CRC could obtain relatively similar accuracy with SRC but with massively reduced time complexity. The experimental results show that our proposed method demonstrates a higher accuracy compared to the traditional method.