Google

Review model errors to identify a recurring mistake pattern, then design input features that address that pattern. These targeted features can reduce avoidable bias, but their effect on variance should be evaluated empirically. If variance increases, regularization can help control it.

Using Error Review to Guide Feature Design

What should you do if adding new features raises variance?

Do Newly Added Inputs Affect Only One Source of Error?

Improving Features from _____ Findings

Roles in Adding New Features

Steps for Adding a New Feature After Error Review

Question:
What is the main theoretical concern when you add new input features after doing error analysis, and what practical step can you take if that concern appears in experiments?

Sample answer:
Adding more features can theoretically make the model more flexible and therefore increase variance. That means the model may fit the training data more closely but become less stable on new examples. If experiments show that the added features are helping but also raising the validation error gap, a practical fix is to use regularization to control complexity. In many cases, regularization offsets the extra variance so the new features can still improve performance overall.

Key points:
- Extra features can raise variance in theory.
- The practical response is regularization.
- Regularization helps control overfitting and can reduce the variance increase.

Rubric:
A strong response will identify higher variance as the theoretical risk and name regularization as the practical method used to address it.

Why Extra Features Can Backfire Temporarily

Case context:
A machine learning engineer reviews errors in a bank card fraud detector and finds that it misses a group of transactions with an unusual combination of merchant category and time-of-day patterns. To address that specific weakness, the engineer adds new engineered input features that highlight those patterns. The training error falls, but validation performance now shows a noticeable rise in variance.

Question:
According to the standard approach for handling this tradeoff after improving bias with new features, what should the engineer do next to address the increased variance?

Sample answer:
The engineer should apply regularization, which is the usual next step for reducing the variance that can appear after adding features intended to lower avoidable bias.

Key points:
- The added features were designed to reduce avoidable bias in a targeted error category.
- The side effect observed was an increase in variance.
- Regularization is the appropriate technique for counteracting that variance increase.

Rubric:
The answer must correctly identify regularization as the needed intervention to address the rise in variance.

Reducing Variance After Adding Diagnostic Features

Question:
After examining the model’s mistakes, you decide to add new input variables. What should those variables be chosen to do?

Sample answer:
They should be built to cut down the specific pattern of mistakes that the review uncovered, rather than to improve all errors equally.

Key points:
- Use the review to identify a recurring mistake pattern.
- Add features that reduce that particular kind of error.

Rubric:
Must state that the added variables are meant to reduce a specific error pattern or class of mistakes found in the review.

Learn Before

Related