Databricks Machine Learning Professional Question 270
Select 3You are monitoring a deployed machine learning model for potential concept drift and feature drift. Which of the following steps represent a common workflow for identifying and measuring these drifts?
- A
Compare the statistical distributions of input features in training data and incoming data.
- B
Monitor the model's performance over time using metrics like accuracy or F1 score.
- C
Train a secondary model to predict the source (training or incoming data) based on the feature values.
- D
Directly compare the weights of the trained model with a newly trained version on incoming data.
- E
Visualize and analyze the importance of features in the original training data only.
Show answer and explanation
Correct answers: A, B, C
Explanation
Detecting concept drift and feature drift requires comparing distributions of training and incoming data, monitoring model performance over time, and using techniques like secondary models to identify differences. These approaches provide a comprehensive understanding of how the data and model behavior are changing in production.
- A. Correct.
This is a valid step to measure feature drift. By comparing the statistical distributions of input features, you can detect significant changes in their behavior.
- B. Correct.
This is a valid step to measure concept drift. Monitoring key performance metrics over time helps identify when the model's predictive performance starts to degrade.
- C. Correct.
This is a valid step to detect both concept and feature drift. A secondary model can highlight discrepancies between training and incoming data distributions.
- D. Incorrect.
Directly comparing model weights is not a standard practice for identifying drift, as differences in weights alone do not provide a clear indication of drift.
- E. Incorrect.
Focusing only on the training data does not help detect drift in incoming data or changes in distributions over time.