EasyDeepLearn

How do you justify shipping logistic regression over a boosted ensemble that scores better?

medium

Answer

  • Frame it as total cost, not offline metric.
  • If the gap is a fraction of a point of AUC and the decision threshold sits in a region where both models rank the same customers, the business outcome is identical.
  • Logistic regression gives you coefficients a regulator can read, near-zero inference latency, trivial monitoring, and calibrated probabilities out of the box.
  • It also fails predictably under drift instead of in surprising ways.
  • Quantify the gap in the metric the business actually uses, and if it is negligible, the simpler model is the better engineering decision.
Check yourself — multiple choice
  • Always ship the highest-AUC model
  • Translate the gap into business impact; if negligible, interpretability, latency, calibration and predictable failure win
  • Simpler models are never justified
  • Only regulators matter

Compare total cost and business impact rather than the raw offline metric.

#interview-scenarios#interpretability

Practise Supervised Learning

215 interview questions in this topic.

Related questions