New Orleans Police Response Time Prediction

A machine learning project to predict police response times for the New Orleans Police Department using incident data from 2025. This project analyzes emergency response patterns to help optimize resource allocation and improve public safety services.

The data attached ds.csv is a real-world dataset from https://data.gov

Project Overview

This project aims to predict the overall response time for police incidents in New Orleans, calculated as the difference between TimeClosed and TimeCreate. By analyzing historical incident data from 2025, we develop predictive models that can help the New Orleans Police Department better understand response patterns and optimize their operations.

Dataset Information

Data Source: Orleans Parish Communication District (OPCD) - the administrative office of 9-1-1 for the City of New Orleans

Dataset Details:

• Time Period: 2025 incident reports
• Target Variable: Response time (TimeClosed - TimeCreated)
• Total Records: 29,753 incidents
• Features: 21 columns including incident types, priorities, locations, and timestamps

Column Descriptions

Column	Type	Description
NOPD_Item	String	Unique incident identifier
Type	String	Incident type code
TypeText	String	Human-readable incident type
Priority	String	Incident priority level
InitialType	String	Initial incident type code
InitialTypeText	String	Initial incident type description
InitialPriority	String	Initial priority assignment
MapX	Numeric	X-coordinate location
MapY	Numeric	Y-coordinate location
TimeCreate	DateTime	Incident creation timestamp
TimeDispatch	DateTime	Dispatch timestamp
TimeArrive	DateTime	Officer arrival timestamp
TimeClosed	DateTime	Incident closure timestamp
Disposition	String	Incident resolution code
DispositionText	String	Resolution description
SelfInitiated	String	Whether incident was self-initiated
Beat	String	Police beat designation
BLOCK_ADDRESS	String	Incident location
Zip	Numeric	ZIP code
PoliceDistrict	Numeric	Police district number
Location	String	Geographic coordinates

Getting Started

Prerequisites

• Python 3.8+
• Jupyter Notebook or JupyterLab
• Required libraries (see requirements section)

Installation

1. Clone the repository:

git clone https://github.com/ysandansing/responseTimePrediction.git
cd responseTimePrediction

2. Create a virtual environment (Optional):

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

or use conda

conda create -n myEnv python=3.8
conda activate myEnv

3. Install required dependencies:

pip install -r requirements.txt

Methodology

Model Development

Baseline Models
Three regression approaches were implemented to establish performance benchmarks:

• Linear Regression: Provided foundational insights into linear relationships between features and response times
• Lasso Regression (L1): Automated feature selection through coefficient zeroing, handling high-dimensional data
• Ridge Regression (L2): Maintained all features while preventing overfitting through regularization

Neural Network Architecture
A 3-layer feedforward network with PyTorch implementation:

- Input (64) → Hidden 1 (128, ReLU, Dropout 0.2) → Hidden 2 (64, ReLU) → Output (1)
- Trained with AdamW optimizer (cyclical LR: 0.001-0.0001)  
- Incorporated batch normalization and early stopping (patience=3)

Model Comparison

Model	Training Time	Validation MSE	Validation MAE	R² Score
Linear Regression	0.09s	0.872	0.621	0.412
Lasso Regression	1.25s	0.763	0.587	0.486
Ridge Regression	0.03s	0.758	0.584	0.491
Neural Network	22m7s	0.682	0.512	0.573

Key Findings:

1. Neural networks achieved 27% lower MAE than best linear model (0.512 vs 0.584 minutes)
2. Lasso/Ridge showed comparable performance despite different regularization approaches
3. Training time scaled 1000x from linear (0.09s) to NN (22m) models
4. Log-transform reduced target variable skewness (right-skew σ from 4.2 → 0.8)