Multimodal Deep Learning Crime Prediction Using Tweets

Abstract

Crime prevention relies on crime prediction as a crucial method to determine the most effective patrol strategy for law enforcement agencies. Various approaches and solutions have been utilized to predict criminal activity. Nonetheless the environment and nature of information for crime prediction is constantly changing. The utilization of social media for sharing information and ideas has experienced a significant surge. Twitter, in particular, is regarded as a valuable platform for gathe ring public sentiments, emotions, perspectives, and feedback. In this regard, techniques for analyzing the sentiment of tweets on Twitter have been developed to ascertain whether the textual content conveys a positive or negative viewpoint on crime incident. Data fusion is a significant technique to integrate the information from crime and tweets data source. Therefore, this study aims to leverage semantic knowledge learned in the text domain and historical crime data, and transfer them into a model trained crime prediction. We applied data fusion technique to ConvBiLSTM model to extract independent vector from tweet and crime modalities and fuse them into a single representation that captures the information from all modalities. This study involved collecting and conducting experiments using two datasets. The first dataset consisted of crime incident data obtained from the Chicago police department, specifically covering the period between September 1 and September 30, 2019. The second dataset comprised tweets containing crime-related terminology specific to Chicago. To evaluate the performance of our model, we benchmarked with latest crime prediction models, including SVM, Logistic Regression, NAHC, DNN with feature-level data fusion, CrimeTelescope, ANN+BERT, and BERT-based crime prediction models. The experimental result showed that our ConvBiLSTM model using multimodal data fusion demonstrates superior performance compared to other traditional deep-learning and BERT models with an accuracy of 97.75%.