Time-Aware CrimeVec: Integrating Time2Vec for Spatio-Temporal Crime Embedding and Arrest Prediction

Crime analysis in urban areas is a critical component of community security and effective operation of law enforcement in cities today. Statistical summaries are used in most conventional methods of crime analysis. The episodes of crime are handled separately and do not indicate complex spatial and time associations among different types of crimes. Conventional data-driven approaches, e.g. criminal embedding techniques using Word2Vec, consider only the co-occurrence patterns, although generally not the temporal dynamics of the data, and as a result, are limited in prediction power. In order to address these limitations, in this paper, we present a novel Time-Aware CrimeVec (TA-CrimeVec) system that takes into consideration the temporal encoding and spatial-temporal crime embedding. The model is trained using Chicago Crime Dataset and uses Time2Vec to encode timestamps using linear and periodic functions. The criminal incidents are transformed to time-sensitive tokens and fed through a Skip-Gram based Word2Vec model to create hybrid embeddings that reflect crime type, time trends and spatial closeness. We then evaluate the learnt embeddings in predicting arrests with ML models like the RF and SVM. The results of the experiment demonstrate that the proposed method can attain a higher level of representation and prediction results, whereas the accuracy of the Random Forest is approximately 78.6 percent. The proposed solution is quite appropriate in capturing what happens when. It gives more information on the crime pattern and makes the decision making more informed in urban safety systems.