Databricks Machine Learning Professional Question 231
Select 2A company is using a machine learning model to predict customer churn based on customer behavior data. Over the past year, the company has introduced new subscription plans and shifted its marketing strategy to target a different demographic. Which of the following scenarios are likely to result from these changes?
- A
Feature drift due to changes in customer behavior patterns associated with the new subscription plans.
- B
Label drift due to changes in the definition of customer churn influenced by the new target demographic.
- C
Feature drift due to the addition of new features to the dataset that represent the marketing strategy.
- D
Label drift due to inconsistencies in how customer churn is recorded in the dataset over time.
- E
No feature drift or label drift since the machine learning model was trained on historical data.
Show answer and explanation
Correct answers: A, B
Explanation
Feature drift occurs when the distribution of input features changes over time, such as when customer behavior patterns shift due to business changes like new subscription plans. Label drift occurs when the distribution or definition of the target variable changes, such as when the company targets a new demographic that may have different churn behavior. Recognizing these scenarios is critical for maintaining model performance.
- A. Correct.
Correct. Changes in customer behavior patterns due to new subscription plans can alter the distribution of input features, leading to feature drift.
- B. Correct.
Correct. If the target demographic has shifted significantly, the definition of customer churn may also evolve, potentially causing label drift.
- C. Incorrect.
Incorrect. Adding new features to the dataset does not in itself constitute feature drift unless the distribution of existing features changes.
- D. Incorrect.
Incorrect. Label drift is related to changes in the meaning or distribution of the target variable, not inconsistencies in recording data.
- E. Incorrect.
Incorrect. Training on historical data does not prevent feature drift or label drift if the underlying data distribution changes over time.