Don't Waste Data
summary
Historical USPS event logs get reduced to a transit-time target (earliest vs latest event) and fed into a scikit-learn random forest regression trained on partner and destination zip features; using ~500k data points (50% of available) the model cuts average error from 1.43 to 0.86 days. The model reveals nonintuitive routing choices (farther print partners sometimes faster due to seasonal congestion), is built as a prototype routing predictor during a two-day hackathon, and is not yet live but shows clear gains.
AI-written summary · about
related talks
42mUsing Data to Understand a New Transportation Reality
26mDelivering Delivery Time Prediction
42mShould we stop using distance in our location-based data recommendation models?
28mContainer Tracking for Bosch Tools Supply Chain Operations
15mThe turf wars on data - can be a matter of life and death
34mMachine Learning to know if you R coming
16mTidy Transit: Real Life Data Modeling for Public Transportation
28m"Where, When, Why, and How: Directi