Streets to Avoid Recommender: High-Risk Street Prediction Model

Currently only available for Washington, USA

Baisakhi Sarkar

Master of Science in Data Science at University of Washington, Seattle

Contact Information: [email protected]/[email protected]

Objective

The goal of this project is to predict high-risk times on specific streets and recommend streets to avoid based on accident likelihood. This model identifies times when certain streets are more prone to accidents, helping drivers, especially vulnerable groups like the elderly, to avoid these areas during high-risk periods. Traffic management agencies can also use this model to optimize resource allocation and proactively issue warnings.

Approach

1. Feature Selection:

   • Key Predictive Features: Street, Start_Time, Severity, Visibility(miles), Precipitation(in) and Weather_Condition.
   • Data Transformation:
     • Time-Based Features: Extract details like hour of the day, day of the week, and month from Start_Time to capture patterns in accident occurrences.
     • Accident Severity: This feature helps highlight higher-risk times with more severe accidents.
     • Environmental Conditions: Visibility(miles), Precipitation(in) and Weather_Condition provide essential information, as weather has a major impact on accident risk.

2. Modeling Approach:

   • Data Preprocessing: Encode categorical variables (e.g., Weather_Condition, Street, City) and standardize continuous features like Visibility.
   • Classification Models:
     • Random Forest: A robust tree-based model effective for handling complex relationships. (Preferred)
     • Gradient Boosting: An ensemble approach that builds trees sequentially, which often excels with complex interactions.
   • Rule-Based Models: Simple rule-based models (e.g., "avoid icy streets at night") can complement machine learning predictions, adding interpretability to the model’s output.

3. Training and Evaluation:

   • Evaluation Metrics: Metrics like ROC-AUC, Precision, and Recall are essential as accident prediction involves imbalanced classes (high-risk vs. low-risk).
   • Hyperparameter Tuning: Use Grid Search or Randomized Search to fine-tune the model, focusing on accuracy and recall for high-risk predictions.

Potential Models

1. Random Forest: Handles non-linear relationships and interacts well with complex data.
2. Gradient Boosting: Known for strong performance on complex data but requires tuning.
3. Rule-Based Models: These straightforward, interpretable models can act as baselines or complement more advanced models.

Outcome

The model will recommend streets to avoid at specific times based on accident likelihood. The output will benefit:

• Traffic Management: Agencies can allocate resources, set up warnings, or divert traffic based on the model’s recommendations. By identifying high-risk times and locations, traffic control can be optimized for better safety and efficiency.

• Driver Safety: Helps drivers, especially those more vulnerable to accidents (e.g., elderly drivers), avoid high-risk streets during hazardous periods. Personalized recommendations based on time, weather, and road conditions offer an extra layer of security.

• Seasonal and Time-Based Analysis: Provides insights into seasonal accident trends (e.g., winter evenings or summer weekends with higher accident rates) to help traffic agencies plan targeted safety campaigns and allocate resources more effectively.

• Infrastructure and Maintenance Planning: Identifies roads with high accident likelihood over time, signaling the need for road maintenance, better lighting, or additional traffic signals. This data allows urban planners to prioritize infrastructure upgrades for high-risk locations.

• Personalized Driver Recommendations: Offers customized safety recommendations for drivers, especially those with specific needs (e.g., elderly drivers). This can suggest safer routes based on individual risk factors like visibility and road type.

• Enhanced Emergency Response: Enables pre-positioning of emergency services closer to high-risk areas during peak times, reducing response times in case of accidents.

• Pedestrian and Cyclist Safety: Helps identify zones with higher risks for non-motorized road users, allowing planners to design safer pedestrian crossings or cyclist lanes and consider time-based road restrictions in high-traffic areas.

These outputs make the Streets to Avoid Recommender a versatile tool for improving road safety, guiding urban planning, and supporting data-driven decision-making across various sectors.

Future Work

• Integration with Google Weather API: Incorporate live weather conditions to enhance prediction accuracy.

  • Automated Data Entry: Automatically fill fields like weather and temperature with real-time conditions, reducing manual input and improving ease of use.
  • Location-Based Risk Assessment: Real-time weather updates will make predictions more relevant based on the user’s location.
  • Real-Time Alerts: Integration with navigation apps (e.g., Google Maps, Waze) can provide real-time alerts for high-risk areas, helping drivers adjust their routes dynamically based on live conditions.
  • Weather-Dependent Risk Notifications: By coupling predictions with live weather data, the model can send dynamic risk alerts when adverse weather conditions increase accident risks, enabling proactive adjustments to travel plans.

• Additional Data Sources: Integrate data from local traffic agencies or public road condition APIs for more accurate risk assessment.

This future work will make the Streets to Avoid Recommender more user-friendly and dynamic, providing location-based recommendations and enhancing road safety insights.