Outdated
patterns.
Training data reflects courier habits that have since changed.
Predict the route.
Update the ETA.
ShanghaiHangzhouChongqingYANTAIJilin.
Train for the
most accurate ETA.
Learn from past deliveries to reduce arrival-time error.
One history.
Three different jobs.
- Before Sep 1
Train.
Earlier checkpoints teach the model its patterns.
- Sep 1–15
Validate.
A later slice helps choose when to stop training.
- Sep 16–30
Evaluate.
Later checkpoints measure ETA error, without fitting to them.
The Yantai development split, not a live training process. The changing delivery reflections are decorative stock imagery, not examples from these groups. Ribbon widths reflect saved counts: 800,000 fitting queries, 181,388 early-stopping queries, and 155,296 evaluation queries. These are checkpoint queries, not unique parcels. Only the fitting band continues into the model illustration; the later groups remain separate. The fitting sample is capped at 800,000. Training uses events and outcomes before September 1. Early stopping uses September 1–15 events with outcomes known before September 16. September 16–30 is development evaluation; no October test is shown.
LightGBM learns from past checkpoints, combining trees to refine the ETA.
Outdated
Frozen
updates.
The courier keeps delivering, but the log stops updating.
Wrong
values.
A delivery has the wrong timestamp or location.
Missing
events.
A parcel is delivered, but the event isn’t recorded.
Duplicate
records.
The same delivery event is recorded more than once.
Data
leakage.
A prediction uses delivery details that were only known later.
One delivery.
More than one story.
Fictional perspectives. Not testimonials.
“Predict my next stop. Don’t choose it for me.”
The courier
“Tell me which scan changed the ETA.”
Dispatch desk
“How many stops come before my parcel?”
“A timestamp doesn’t tell me when the record arrived.”
The data reviewer
“One stop done. Which parcels are still on the van?”
“The ETA changed. What happened?”
The customer