Leveraging transfer learning with deep learning for crime prediction.

Area of Science:

• Computational Criminology
• Artificial Intelligence in Public Safety
• Deep Learning for Predictive Analytics

Background:

• Crime prediction is vital for public safety, with deep learning showing promise.
• Limited crime data and resources hinder the training of advanced deep learning models.
• Existing statistical and deep learning methods require significant data for effective crime forecasting.

Purpose of the Study:

• To address data scarcity in crime prediction by employing the transfer learning paradigm.
• To fine-tune and evaluate various statistical and deep learning models for crime forecasting.
• To propose and validate a novel BiLSTM-based transfer learning architecture for enhanced crime prediction accuracy and efficiency.

Main Methods:

• Fine-tuning of statistical methods: Simple Moving Averages (SMA), Weighted Moving Averages (WMA), Exponential Moving Averages (EMA).
• Fine-tuning of deep learning models: Long Short Term Memory (LSTM), Bi-directional Long Short Term Memory (BiLSTM), and Convolutional Neural Networks and Long Short Term Memory (CNN-LSTM).
• Development and evaluation of a BiLSTM-based transfer learning architecture for transferring crime prediction knowledge across different neighborhoods.

Main Results:

• The proposed transfer learning approach with BiLSTM demonstrated superior performance compared to other methods.
• Achieved significantly low error values in predicting weekly and monthly crime trends across diverse datasets (Chicago, New York, Lahore).
• Demonstrated reduced execution time, indicating computational efficiency for practical law enforcement applications.

Conclusions:

• Transfer learning, particularly with BiLSTM, effectively overcomes data limitations in crime prediction.
• The proposed model enhances the accuracy and efficiency of crime forecasting.
• This approach offers a valuable tool for law enforcement agencies to improve crime control and prevention strategies.