CrimePredict+

A study comparing 3D CNN, RNN, and CNN+RNN models for crime hotspot prediction.

  • Master Project by Zhen Lu

CrimePredict+ is a focused study that compares the performance of three advanced deep learning models—3D CNN, RNN, and CNN+RNN—on call-for-service data to predict future crime hotspots. The study determines the strengths and weaknesses of each architecture in predicting crime hotspots, providing critical insights for improving predictive accuracy and scalability in crime prevention applications.

Objectives

  1. Data Transformation: Develop methods to convert call-for-service data into spatio-temporal formats suitable for each model.
  2. Model Comparison: Compare the performance of 3D CNN, RNN, and CNN+RNN architectures on crime hotspot prediction tasks.
  3. Model Evaluation: Train, test, and evaluate models to measure accuracy, precision, recall, and F1-score.

Technical Stack

  • Dataset: Call-for-service data from Portland, Oregon, formatted into spatio-temporal structures.
  • Machine Learning Models:
    • 3D CNNs: Analyze spatial and temporal relationships using 3D convolutional kernels.
    • RNNs (LSTM): Capture temporal dependencies in sequential data.
    • CNN+RNN: Combine CNN’s spatial feature extraction with RNN’s temporal analysis.
  • Frameworks and Libraries:
    • TensorFlow for model implementation.
    • Pandas and Numpy for data transformation.
    • Matplotlib and Seaborn for result visualization.
  • Evaluation Metrics:
    • Accuracy, precision, recall, and F1-score to assess model performance.