Build a usable dataset
Combine NYC collision records with weather and speed-limit data, then prepare the features and address missing values.
Predictive modelling · 2021
Understanding serious collisions in New York City.
Combining collision records, weather and speed-limit data to examine road safety patterns and experiment with models of accident severity.
Public datasets seldom fit neatly together.
Road accidents happen in a wider context. Location, timing, traffic conditions and weather may all matter, but the relevant information is spread across multiple sources.
The project combined descriptive analysis with predictive modelling, seeking to understand accident-severity patterns and present the findings in a format people could explore.
Integrate data. Map patterns. Test models.
The work linked exploratory analysis with interactive data storytelling and traditional supervised machine learning.
Make patterns visible
Investigate temporal trends, geographic hotspots and contributing factors using plotting, Folium maps and Bokeh dashboards.
Compare classical classifiers
Experiment with logistic regression, decision trees and random forests while considering class balance and feature importance.
What the project delivered.
Explorable patterns
The project produced visual analyses and interactive dashboards around collision patterns and accident severity.
A supervised modelling approach
It experimented with classification methods to assess accident severity using features from multiple data sources.
Features worth investigating
The original write-up discusses contextual variables such as location, vehicle type and weather as potentially relevant signals.
Good analysis is more than a prediction.
The project demonstrates the value of combining data engineering, modelling and visual storytelling. A model score is only one part of making a dataset useful.
Presenting limitations and helping readers explore the underlying patterns can be just as important as selecting a particular algorithm.
See the original visual story.
The original project link contains the deeper interactive exploration, technical details and supporting analysis.