Crime Hot Spot Forecasting

Integrated transposed convolutions and deep recurrent networks to capture complex spatio-temporal patterns in crime data and generate accurate forecasts of crime hotspots.

The work on this project has published in ICBK 2017. Take a look at “Crime hot spot forecasting: A recurrent model with spatial and temporal information” for more details.

Objective

In recent years, place-based crime prediction has garnered significant attention from researchers. Numerous studies have been conducted on this topic, and some of them have found that in specific areas, there are stable hot spots and flow hot spots that exist simultaneously. Moreover, the measurement of crime hot spots becomes more dynamic and unstable as the time and space scales decrease. Therefore, capturing the temporal evolution of spatial crime patterns has become crucial in developing effective and efficient crime prevention strategies.

Proposed Model

Deep recurrent neural networks (RNN) have shown remarkable performance in numerous sequence prediction problems, such as machine translation, contextual parsing, image captioning, and video description. These networks are specially designed to process sequences and capture the temporal evolution of the objects within them. Additionally, some variants like the Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) have demonstrated an ability to handle long-term dependencies particularly well. Motivated by this, we built a deep network to forecast crime hot spots by simulating spatial patterns’ temporal evolution via embedding spatial information into recurrent architecture. We believe that capturing these spatial patterns’ temporal evolution can enhance the effectiveness and efficiency of crime prevention strategies. To evaluate the effectiveness of our model, we used call-for-service data provided by the Portland, Oregon Police Bureau (PPB) over a five-year period from March 2012 through December 2016. Our experiment results indicate that our proposed model outperforms several classical machine learning approaches and alternative neural network architectures in forecasting crime hot spots.