avnisinghal001/narisafe-risk-awareness-model

NariSafe Risk Awareness Model

NariSafe is a public-data-based women safety risk-awareness prototype using NCRB/OpenCity crime statistics, OpenStreetMap urban infrastructure features and engineered contextual features.

The model predicts a contextual risk-awareness level:

• low
• medium
• high

Final Model

The uploaded model is:

No-location HistGradientBoostingClassifier with high-risk threshold tuning

Direct location identity features were removed from model input:

• city
• latitude
• longitude

These features are used only for lookup/display in the app, not for prediction.

Evaluation

Primary benchmark: group-city split, where test cities are unseen during training.

Final threshold-tuned result:

• Test accuracy: 0.8837
• Test macro F1: 0.8163
• Test weighted F1: 0.8832
• High-class precision: 0.8015
• High-class recall: 0.5769
• High-class F1: 0.6709

Files

• best_no_location_threshold_model.joblib — final trained sklearn model bundle
• feature_config_no_location.json — feature configuration used by the model
• high_threshold_tuning.csv — threshold tuning results
• baseline_model_report.md — baseline evaluation report
• inference_example.py — sample inference logic

Important Limitation

This model does not predict guaranteed real-world crime.

The target label risk_level is a rule-based proxy label created for prototype training because real incident-level ground-truth labels were not available.

Predictions should be interpreted as risk-awareness levels or model confidence, not real-world probability of crime.

Ethical Use

This model is intended for educational/demo use and public-data-based awareness only.

Do not use it as:

• an emergency decision system
• a policing tool
• an official safety advisory
• an individual risk scoring system
• a real-time crime prediction system