<![CDATA[Crime analysis using various types of data, such as video surveillance, crime reports, and criminal networks, has been widely investigated in digital criminology. Most of the available data are unlabelled. In this work, we introduce a self-supervised learning framework for multimodal criminology, which enables the fully automatic learning of effective features for unlabelled video, text, and graph datasets and the completion of crime analysis tasks, including anomaly detection, crime report classification, and high-risk node prediction via contrastive learning, masked prediction, and graph self-supervised learning. The experimental results show that our SSL model learns high-quality features and achieves better performance than its supervised counterpart and baseline models. Unlike traditional deep learning-based models that require large amounts of labeled data, our proposed SSL model is label-efficient, scalable, and robust to artificial or anonymous data. Our work aims to develop an AI-based multimodal self-supervised learning approach for efficient, accurate, reliable, and safe crime analysis]]>
Copyrights © 2026
