Hybrid of deep learning and exponential smoothing for enhancing crime forecasting accuracy

Area of Science:

• Urban planning and criminology
• Data science and machine learning
• Public safety and law enforcement

Background:

• Urbanization presents significant challenges for law enforcement in maintaining public safety.
• Crime prevention strategies are often budget-intensive, yet crime prediction remains underdeveloped.
• Existing machine learning and time series methods for crime forecasting face accuracy limitations due to data deficiencies.

Purpose of the Study:

• To develop an accurate crime forecasting model using spatiotemporal data.
• To address the limitations of current crime prediction techniques.
• To propose a hybrid Bidirectional Long Short Term Memory (Bi-LSTM) and Exponential Smoothing (ES) model for enhanced crime forecasting.

Main Methods:

• Utilized a hybrid approach combining Bidirectional Long Short Term Memory (Bi-LSTM) and Exponential Smoothing (ES).
• Evaluated the model's performance on New York City crime data spanning from 2010 to 2017.
• Compared the proposed hybrid model against the Seasonal Autoregressive Integrated Moving Averages (SARIMA) benchmark.

Main Results:

• The Bi-LSTM and ES hybrid model demonstrated superior performance compared to SARIMA.
• Achieved significantly lower error metrics, including Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE).
• Specific error metrics: MAPE (0.3433), RMSE (13.104), MAE (9.837).

Conclusions:

• The proposed hybrid Bi-LSTM and ES model offers a more accurate method for crime forecasting.
• This technique can empower law enforcement agencies with better tools for crime prevention and control.
• Improved crime pattern prediction facilitates proactive public safety measures.