Databricks Machine Learning Professional Question 244
Select 3You are monitoring a machine learning model deployed in production that predicts customer churn. Recently, you noticed a significant drift in one of the categorical features, 'Subscription_Type', where the distribution of its unique values has changed. Which of the following approaches can help you identify and understand the drift in this categorical feature?
- A
Calculate the mode of the 'Subscription_Type' feature in both training and production data and compare them.
- B
Identify the unique values in the 'Subscription_Type' feature from both training and production data and compare the sets.
- C
Replace missing values in the 'Subscription_Type' feature with a constant value to eliminate drift completely.
- D
Compare the proportions of missing values in the 'Subscription_Type' feature between training and production data.
- E
Check the correlation between the 'Subscription_Type' feature and the target variable in production data.
Show answer and explanation
Correct answers: A, B, D
Explanation
Categorical feature drift can be identified through methods that analyze the distribution and characteristics of the feature, such as comparing modes, unique values, and missing value proportions across training and production datasets. These approaches provide insights into how the feature has changed over time, helping to diagnose and address the drift.
- A. Correct.
Calculating and comparing the mode of the feature in training and production data can highlight changes in the most frequent category, which is a potential indicator of drift.
- B. Correct.
Comparing the unique values in training and production data can reveal if new categories have emerged or if certain categories are no longer present, which is critical for understanding drift.
- C. Incorrect.
Replacing missing values with a constant does not address the root cause of drift and does not provide insights into the nature of the drift.
- D. Correct.
Comparing the proportions of missing values between training and production data helps identify if missingness has increased, which may contribute to drift in the feature.
- E. Incorrect.
Checking the correlation with the target variable is not directly relevant for detecting or understanding categorical feature drift, as it focuses on the relationship with the target rather than the feature distribution itself.