Research map / Learning theory and algorithms / Prediction and decision losses
Data-Driven Structural Fire Risk Prediction for City Properties
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Applies XGBoost (with missing-value handling and positive class weighting) to historical city property and fire-incident data from a South Dakota fire department to rank properties by structural fire risk for inspection prioritization.
Problem
Fire departments have limited inspection resources and allocate them by intuition and simple rules rather than data-driven fire risk estimates across city properties.
Approach
Frame fire inspection prioritization as ranking properties by predicted fire probability learned from property, inspection, violation and crime records, using XGBoost's native handling of missing values and class weighting rather than imputation or synthetic oversampling.
Key results
- XGBoost on full data with missing values reaches test AUC 0.757 (320 test fires); a weight of 1 places 16 fires in the top 100 versus 1 for random ranking.
- On the reduced dataset without missing values (4,811 properties, 172 fires), Random Forest achieves AUC 0.699 with scaling, outperforming XGBoost (0.603) and AdaBoost (0.656).
- Increasing the positive class scaling weight lowered validation AUC in the full-data XGBoost setting (0.772 at weight 1 to 0.715 at weight 100).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use XGBoost rankings instead of a random ranking | XGBoost on the full dataset · South Dakota properties, top 100 ranked properties | properties with fire incidents: 1 -> 16 | Helped |
| Increase the positive scaling weight from 1 to 50 (ablation) | XGBoost · South Dakota properties, five-fold validation | AUC: 0.772 -> 0.742 | Hurt |
| Increase the positive scaling weight from 1 to 73 (ablation) | XGBoost · South Dakota properties, five-fold validation | AUC: 0.772 -> 0.724 | Hurt |
| Increase the positive scaling weight from 1 to 100 (ablation) | XGBoost · South Dakota properties, five-fold validation | AUC: 0.772 -> 0.715 | Hurt |
| Remove properties with missing values (ablation) | XGBoost · South Dakota properties | AUC: 0.757 -> 0.598 | Hurt |
| Scale positive and negative classes to equalize them (ablation) | XGBoost on the reduced dataset · South Dakota properties, reduced dataset | AUC: 0.598 -> 0.603 | Helped |
| Scale positive and negative classes to equalize them (ablation) | ADABoost on the reduced dataset · South Dakota properties, reduced dataset | AUC: 0.625 -> 0.656 | Helped |
| Scale positive and negative classes to equalize them (ablation) | Random Forest on the reduced dataset · South Dakota properties, reduced dataset | AUC: 0.693 -> 0.699 | Helped |
Surprising findings
- Upweighting the positive (fire) class decreased AUC for XGBoost on the full dataset, contrary to the usual expectation.
- Imputation with KNN or linear regression before Random Forest/AdaBoost did not match XGBoost's native missing-value handling on the full dataset.
Limitations
- Private, non-public data from a single South Dakota department limits reproducibility and comparison.
- Reduced dataset after removing missing-value properties is much smaller and more imbalanced, hurting performance.
- Feature set was not extensively explored; interplay between positive weighting and XGBoost missing-value handling is unresolved.
Experimental setup
Datasets: South Dakota fire department property records fire incident data (structural fires, fire code 111) code case violations crime data utility disconnections rental registrations fire inspections parcel data
Benchmarks: held-out test split of South Dakota city property data
Metrics: AUC cumulative fire count curves
Baselines: AdaBoost Random Forest random ranking KNN imputation with XGBoost/AdaBoost/Random Forest linear regression imputation
Builds on: XGBoost SMOTE
Closest papers by meaning
- FireScope: Wildfire Risk Raster Prediction With a Chain-of-Thought Oracle (CVPR 2026 · similarity 0.82)
- DetectiumFire: A Comprehensive Multi-modal Dataset Bridging Vision and Language for Fire Understanding (NeurIPS 2025 · similarity 0.79)
- Active Geospatial Search for Efficient Tenant Eviction Outreach (AAAI 2025 · similarity 0.79)
- UrbanDataLayer: A Unified Data Pipeline for Urban Science (NeurIPS 2024 · similarity 0.78)
- Sim2Real-Fire: A Multi-modal Simulation Dataset for Forecast and Backtracking of Real-world Forest Fire (NeurIPS 2024 · similarity 0.78)
- On the Efficiency of ERM in Feature Learning (NeurIPS 2024 · similarity 0.77)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.