Modeling Daily Crime Events Prediction Using Seq2Seq Architecture

Abstract

Early prediction of the crime occurrence reduces its impact. Several studies have been conducted to forecast crimes. However, these studies are not highly accurate, particularly in short-term forecasting such as over one week. To respond to this, we examine sequence to sequence (Seq2Seq) based encoder-decoder LSTM model using two real-world crime datasets of Brisbane and Chicago, extracted from the open data portal, to make one week ahead of total daily crime forecasting. We have built an ARIMA statistical model and three machine learning-based regression models that differ in their architecture, namely, simple RNN, LSTM, and Conv1D with a novel approach of walk-forward validation. Using a grid search strategy, the hyperparameters of the models are optimized. The obtained results demonstrate that the proposed Seq2Seq model is highly effective, if not superior, compared to its counterparts and other algorithms. This proposed model achieves state-of-the-art results with a relatively Root Mean Squared Error (RMSE) of 0.43 and 0.86 on both datasets, respectively.