A manufacturing firm is choosing a predictive model for equipment-failure alerts. Two candidates have completed evaluation:
Model A
Regularized logistic regression that uses the 12 raw sensor variables.
10-fold cross-validated AUROC = 0.912 ± 0.006
Model B
Gradient-boosted decision trees built from 300 engineered features.
10-fold cross-validated AUROC = 0.915 ± 0.009
Both models meet the minimum business requirement of AUROC ≥ 0.90 on an unseen test set, and the production environment has strict limits on CPU time for daily retraining and scoring.
According to Occam's razor (law of parsimony), which deployment decision is most appropriate?
Deploy Model A; its slightly lower AUROC still meets requirements, and its simpler hypothesis avoids unnecessary complexity.
Create an ensemble of Models A and B so that their predictions can be combined to maximize accuracy.
Collect more data and build an even more complex deep-learning model before moving to production.
Deploy Model B because its marginally higher AUROC proves superior generalization, regardless of added complexity.
Occam's razor recommends preferring the hypothesis that achieves the goal with the fewest unnecessary assumptions. Here, both models satisfy the business performance target, and the 0.003 AUROC difference is within the variation seen across folds. Model A achieves the required generalization while using far fewer features and a simpler algorithm, so it will be faster to retrain, easier to interpret, and less costly to maintain. Selecting the more complex Model B, ensembling the two models, or building an even deeper network would all add complexity without a justified gain in predictive power, contradicting the principle of parsimony.
Ask Bash
Bash is our AI bot, trained to help you pass your exam. AI Generated Content may display inaccurate information, always double-check anything important.
What is AUROC and why is it important in model evaluation?
Open an interactive chat with Bash
What is regularized logistic regression and why is it simpler than gradient-boosted trees?
Open an interactive chat with Bash
What does Occam's razor recommend in machine learning model selection?