Crime Forecasting and Prediction Using Spatio-Temporal Techniques

Conference Presentation

This paper was presented at IJCACI Conference (2025).
You can view the official Letter of Presentation here:

View Certificate / Letter

Overview

This repository contains the implementation of Crime Forecasting and Prediction Using Spatio-Temporal Techniques.
We apply machine learning, deep learning, and graph neural networks (GNNs) to forecast urban crime patterns across time and space using Los Angeles crime data (2020–2025).

---

Methodology

Dataset and Preprocessing

This study uses a comprehensive dataset of over 1 million crime reports from Los Angeles (2020–2025), sourced from the City of Los Angeles Open Data Portal.
The dataset was chosen because it captures post-COVID crime dynamics, reflecting social behavior shifts, law enforcement compliance, and economic conditions.

Preprocessing Steps:

• Datetime Conversion – Converted DATE OCC and TIME OCC into Python datetime format, extracting hour, day, month, year, weekday for temporal granularity.
• Time Binning – Crimes grouped into 6 daily intervals:
  • Late Night (12am–3am)
  • Early Morning (3am–6am)
  • Morning (6am–12pm)
  • Afternoon (12pm–5pm)
  • Evening (5pm–9pm)
  • Night (9pm–12am)
• Additional Features –
  • Day of Week
  • Weekend Flag
  • Public Holiday Indicator (via holidays Python library)
• Crime Category Consolidation – Reduced 68 unique crime descriptions into 17 higher-level categories for clearer pattern recognition (e.g., burglary, theft, assault grouped).
• Scaling/Normalization – For deep learning models (LSTM), features were normalized with MinMaxScaler.

---

Models Implemented

Classification Models (Crime Solvability)

Objective: Predict whether a reported crime will be resolved or unsolved based on historical features.

• Decision Tree (DT): Baseline model → accuracy = 86%
• K-Nearest Neighbors (KNN): Captured neighborhood similarities → accuracy = 90%
• Random Forest (RF):
  • RF (10 trees) → 91%
  • RF (20 trees) → 92% (Best)

Key Takeaway: Random Forest (20 trees) was most effective due to its ability to capture complex decision boundaries in noisy urban crime data.

---

2 Temporal Forecasting Models

Objective: Forecast crime frequency trends over short- and long-term horizons.

• ARIMA + Prophet:

  • ARIMA captures short-term autoregressive patterns
  • Prophet incorporates seasonality, trend changes, holidays
  • Hybrid improved both short and long forecasts

• Prophet + ARIMA:

  • Prophet first captures long-term seasonalities
  • ARIMA refines predictions with short-term fluctuations
  • Outperformed ARIMA-first hybrid

• RNN + LSTM:

  • Bidirectional LSTM with look-back window of 12 weeks
  • Trained for daily, weekly, monthly horizons
  • Used dropout + early stopping to reduce overfitting
  • Best RMSEs:
    • Daily → 0.0585
    • Weekly → 0.06275
    • Monthly → 0.1339

Key Takeaway: RNN+LSTM significantly outperformed classical models, especially in capturing non-linear sequential patterns.

---

Spatial Forecasting Models (Graph Neural Networks)

Objective: Model spatial spillover effects across 21 LA regions. Each region is represented as a node, with edges defined by geographic proximity.

• ST-GCN (Spatio-Temporal Graph Convolutional Network):

  • Combines GCN layers (spatial) + LSTM (temporal)
  • Best performance → RMSE: 244.78

• GAT (Graph Attention Network):

  • Applies attention mechanism to weigh neighbors dynamically
  • RMSE = 251.59 (slightly worse due to noise sensitivity)

• STA-GNN (Spatio-Temporal Attention GNN):

  • Dual attention → spatial + temporal
  • RMSE = 252.88 (did not outperform ST-GCN)

Key Takeaway: ST-GCN proved most effective, striking the best balance between spatial awareness and temporal forecasting.

---

Results

Task	Best Model	Metric	Score
Crime Solvability	RandomForest	Accuracy	92%
Temporal Forecasting	RNN+LSTM	RMSE (daily)	0.0585
Spatial Forecasting	ST-GCN	RMSE	244.78

---