Initial

Initial Problem Statement

The Paterson, NJ Fire Department is one of the busiest in the country, serving ~160,000 residents across 8.4 square miles. Each year, it responds to roughly 300 fires, half classified as major (incidents with significant environmental and public safety consequences that require substantial resource deployment). To reduce fire risk, the department conducts safety inspections of schools, businesses, and multi-family residential properties. But understaffing and a daily volume of 10–12 fire incidents have forced inspections to concentrate on properties with active complaints and registered code violations, making the department more reactive than proactive.

This project aims to help the Paterson Fire Department reduce preventable fires and their impact by recommending which buildings to prioritize for safety inspections. We expect this approach to make better use of limited inspection capacity and help the city move from a reactive response program to proactive prevention and mitigation.

Post-Summer Executive Summary of the Report

The problem

The Paterson Fire Department (PFD) has limited inspection capacity and must decide which properties to prioritize for proactive inspection. Because severe fire-code violations are usually only recorded when an inspection occurs, many uninspected properties have unknown conditions. 

Social impact

Better inspection prioritization can help PFD identify hazardous conditions earlier and reduce fire-related harm, including injury, death, displacement, property loss, and risks to first responders. 

Approach

This project developed a data-informed prioritization framework for eligible active, non-demolished three-or-more-family homes and non-life-hazard businesses in Paterson. The system links inspection, violation, fire incident, building, and location data to identify the 100 highest-risk properties for PFD review every month.

Key Findings

The best-performing machine learning model (an XG Boost classifier), when applied to held-out data from 2025-06-01 to 2026-06-01, resulted in identifying 100 properties a month (1200 for the year). Of these 1200, 344 were actually inspected by PFD. 29% of the inspected ones resulted in a severe violation, capturing 66% of all the severe violations PFD found during that 12-month period. This model outperformed the best heuristic baseline, which found only 56% of all the violations, resulting in 18% improvement using the same inspection capacity..

Interpretation

The model appears to capture useful risk signals beyond simple rule-based prioritization. However, properties that were not inspected should not be treated as safe; their conditions remain unknown.

Recommendations

Given these promising results, we recommend PFD refine this system by 1) verifying and extending the definition of severe violations, 2) verifying and improving the building classifications to determine the inspection-eligible cohort, and 3) conducting a field validation to evaluate this system prospectively, specifically the buildings identified that were rarely or never inspected.

Overall, the model should be treated as a decision-support tool to help PFD identify severe fire-code violations earlier, expand visibility, and use limited inspection capacity more systematically.