Classification-Labeled Continuousization and Multi-Domain Spatio-Temporal Fusion for Fine-Grained Urban Crime Prediction
Shuai Zhao, Ruiqiang Liu, Bo Cheng, Daxing Zhao
Abstract
Fine-grained urban crime prediction is of great significance to urban management and public safety. Previous crime prediction work has been done at a relatively coarse time granularity, which may suffer from two issues for fine-grained crime prediction. 1) <i>The zero-inflation problem</i> associated with fine-grained granularity. Crime occurrence is sparse, and when the time granularity becomes finer, it leads to a more sparse prediction label for this problem resulting in the zero inflation problem. 2) <i>Insufficient amount of information</i> involved in crime datasets. When the spatio-temporal granularity becomes smaller, more information from related fields needs to be introduced to extract spatio-temporal features to assist the analysis. To address the first issue, we introduce a classification-labeled continuousization strategy and a weighted loss function for sparse classification problem, making the model more likely to focus on non-zero elements in zero-inflated datasets. For the second issue, we propose a novel deep learning based model, termed attention-based spatio-temporal multi-domain fusion network, which fuses features from multiple datasets in related domains. We evaluate our method on six real-world datasets collected in New York City and experiments on our model show the advantages beyond many competitive baselines.