Compare Linear and Nonlinear Boundaries
The curved boundary looks much better on training deliveries.
Read the full lesson and open its transcript below. A narrated video will be added after production.
What you will learn
The curved boundary looks much better on training deliveries.
A practical question
The curved boundary hugs every training point. The straight one makes a few training errors. Which should the dispatcher use? Evaluate both on the same future weeks or held-out routes.
The useful idea
Keep preprocessing, target, and metrics identical. Compare severe-delay recall, false warnings, speed, and maintenance needs. A nonlinear model is justified when its gains on relevant cases are real and repeatable; fitting the old chart better is not enough.
Work through the example
Inspect trips on which their decisions disagree. If the curved model wins only on the routes used to select its settings, it may be overfit. If it consistently handles an important interaction, that evidence supports the extra complexity. A nonlinear boundary can wrap around unusual clusters and look impressive on the training plot. Ask whether those clusters also appear in later deliveries or in another depot. Compare a linear baseline and a more flexible classifier with the same features and grouped or time-aware split. If gains vanish on new conditions, prefer the simpler rule or gather evidence before deployment. Complexity earns its place when it changes real decisions reliably, not when it makes a training chart prettier. Use this case to test your understanding: Choose between linear and nonlinear classifiers with grouped or time-aware validation. Here is the limitation to watch for: A complicated boundary should earn its complexity with held-out improvement. This supports a careful conclusion: Model selection compares generalization, latency, and maintenance, not just training fit. Ask whether the curved classifier improves the severe misses the dispatcher cares about, or merely corrects harmless training mistakes. A simple model that generalizes to another depot can be preferable even when its training diagram contains more errors.
Limits and next step
Next, we'll look at a very simple neural learner and connect its mistakes to updates in its weights.
Worked example
Choose between linear and nonlinear classifiers with grouped or time-aware validation.
Quick check: test your understanding
Try it yourself: Consider this case: Choose between linear and nonlinear classifiers with grouped or time-aware validation. Explain how you would check the result. In your answer, address this limitation: A complicated boundary should earn its complexity with held-out improvement.
Check the reasoning
Model selection compares generalization, latency, and maintenance, not just training fit. A complicated boundary should earn its complexity with held-out improvement.
What is the main idea behind compare linear and nonlinear boundaries?
Model selection compares generalization, latency, and maintenance, not just training fit.
How is it used in the example?
Choose between linear and nonlinear classifiers with grouped or time-aware validation.
What limitation should be checked?
A complicated boundary should earn its complexity with held-out improvement.
Further reading
Machine Learning — Tom M. Mitchell.
Original ToolsInVision teaching examples. Editorial approach · Suggest a correction
Accessible referenceLesson 52 transcript292 words · Narration ready for the upcoming videoOpen transcript
The curved boundary hugs every training point. The straight one makes a few training errors. Which should the dispatcher use? Evaluate both on the same future weeks or held-out routes.
Keep preprocessing, target, and metrics identical. Compare severe-delay recall, false warnings, speed, and maintenance needs. A nonlinear model is justified when its gains on relevant cases are real and repeatable; fitting the old chart better is not enough.
Inspect trips on which their decisions disagree. If the curved model wins only on the routes used to select its settings, it may be overfit. If it consistently handles an important interaction, that evidence supports the extra complexity. A nonlinear boundary can wrap around unusual clusters and look impressive on the training plot. Ask whether those clusters also appear in later deliveries or in another depot. Compare a linear baseline and a more flexible classifier with the same features and grouped or time-aware split. If gains vanish on new conditions, prefer the simpler rule or gather evidence before deployment. Complexity earns its place when it changes real decisions reliably, not when it makes a training chart prettier. Use this case to test your understanding: Choose between linear and nonlinear classifiers with grouped or time-aware validation. Here is the limitation to watch for: A complicated boundary should earn its complexity with held-out improvement. This supports a careful conclusion: Model selection compares generalization, latency, and maintenance, not just training fit. Ask whether the curved classifier improves the severe misses the dispatcher cares about, or merely corrects harmless training mistakes. A simple model that generalizes to another depot can be preferable even when its training diagram contains more errors.
Next, we'll look at a very simple neural learner and connect its mistakes to updates in its weights.