Interpretability, Speed, Maintenance, and Decision Cost

Unit ID: AMLA-M01-U02 Estimated active time: 25-35 minutes

Classroom explanation

Model choice is not only about score. It is also about use.

A model can be accurate and still be a poor choice if it cannot be explained well enough, takes too long to train, is difficult to maintain, or fails in costly ways.

Interpretability

Some settings need a simple explanation. A regularised linear model or shallow tree may be easier to discuss than a large ensemble. Inspection tools can help, but they do not turn every complex model into a transparent one.

Ask:

• Who needs to understand the model?
• What kind of explanation is needed?
• Is a rough explanation enough, or is a direct rule needed?

Speed

Training time and prediction time both matter. A slow model can be acceptable for offline analysis and unacceptable for an interactive product. A method that is fine on 1,000 rows may become painful on 1,000,000 rows.

Maintenance

Complex pipelines require more care. If performance is only slightly better, a simpler model may be better because it is easier to rerun, monitor, debug, and explain.

Decision cost

Errors do not have equal cost. A false positive and false negative can lead to different actions. If one error type is costly, threshold, calibration, and error analysis matter as much as the algorithm family.

Practice

For the same problem statement from Unit 00, answer:

• Who will use the prediction?
• What action follows the prediction?
• Which error is more costly?
• Does the model need to be explained?
• Is speed important?

Takeaway

The best model is not always the highest-scoring model. It is the model whose evidence, limits, and operating cost fit the decision.